Packet-Level In-Network Semantic Adaptation for Unstable Mobile Emergency Networks
arXiv:2609.29920v1 [cs.NI] 24 Sep 2026
Zhiyuan Ren∗
Tao Zhang
Wenchi Cheng
School of Telecommunications Engineering, Xidian University Xi’an 710071, China [email protected]; [email protected]; [email protected]
September 2026
Abstract Mobile emergency networks can experience independently changing intermediate wireless links on timescales shorter than endpoint feedback can track. When an egress changes after packet emission, feedback affects only later source data, while the on-path node observes the current condition with the affected packet still mutable. This paper presents DINA, a packetlevel in-network semantic adaptation method. An image is divided into self-describing spatial packets carrying coordinates, a current representation identifier, and payload. At each eligible node, an offline-trained frozen selector scores compatible operators, immediately transforms the packet, and forwards it without image reconstruction or cross-packet adaptation state. Later nodes can retain or further compact the packet through the same typed compatibility contract. The receiver places available packets by coordinate, fills missing regions with black, and runs a fixed machine task. We realize DINA in a 24-node UAV environment using XDP and AF_XDP. In the primary forest-fire trace, DINA raises deadline tile coverage from 40.4% to 72.6% and classification accuracy from 77.5% to 95.0% relative to forwarding. In an independently trained RescueNet segmentation case, it raises coverage from 65.6% to 91.7% and foreground mIoU from 0.486 to 0.541. Sufficient- and extreme-capacity profiles expose a no-gain boundary and a common task-failure boundary, respectively.
Keywords: in-network computing; mobile emergency networks; packet processing; task-aware communication; UAV networks; visual sensing
1
Introduction
Mobile emergency networks are assembled when fixed infrastructure is unavailable, damaged, or unable to cover a hazardous area. UAVs, vehicles, and robots act as sensors and temporary relays, carrying task-bearing observations toward a command endpoint. Mobility, obstruction, contention, and changing relay geometry make multi-hop UAV links time-varying [1, 2], while forest measurements show strong environment-specific propagation effects [3]. Different directed links on a temporary route can therefore evolve on different timescales. A packet can leave its source under one condition and reach an intermediate node whose outgoing link has since become the path’s immediate constraint. ∗
Corresponding author: [email protected]
1
Figure 1: DINA in a mobile forest-fire emergency network. A sensing UAV emits large RGB spatial packets (blue) through temporary relays. The highlighted on-path node observes degradation on its exact outgoing link while those packets remain mutable, transforms them into smaller task-related packets (orange/red), and forwards them across the constrained hop. The faint return path denotes endpoint feedback. Under rapid link variation, its observation can become stale before sourceside adaptation takes effect, and it can change only subsequent packets, not the in-flight packet already present at the constrained relay. The evaluated topology and link processes are specified independently. Endpoint and analytics-edge systems adapt encoder or application settings to measured resource conditions and task utility [4, 5, 6]. When the relevant condition belongs to an intermediate egress, applying that control at the source requires an observe–return–act loop. Under rapid variation, the condition can change again before the source’s adapted traffic reaches that hop. The action also applies only to data still held by the source and cannot retroactively change a packet already waiting at the newly constrained relay. Processing after the weak hop is too late to reduce the load already offered to it. As Fig. 1 shows, the relay is the first execution point where the current state of that egress and the affected, still-mutable packet coexist. This paper presents DINA, a packet-level method that uses this node-local execution opportunity. The source divides an image into self-describing spatial packets carrying coordinates, a current representation identifier, and its payload. At each capable on-path node, an offline-trained and frozen lightweight policy observes the current packet and the node’s exact-egress condition, selects an application-registered operator compatible with the packet’s current representation, transforms the packet, and forwards it immediately. The next packet starts from its own inputs; no image reconstruction or cross-packet adaptation state is involved. Current-representation typing also supports continued processing along a multi-hop path. A downstream node consumes the representation currently carried by the packet and selects only operators that accept that type. The receiver decodes available packets by coordinate, fills missing 2
regions with black, and runs a fixed machine task on the realized canvas. Semantic benefit is therefore measured by the receiver’s task outcome, alongside separate network measurements of delivered coordinates, bytes, queues, and drops. We realize DINA in a configurable 24-node UAV environment using XDP and AF_XDP. The online worker reads a versioned exact-egress observation, evaluates a quantized integer selector, and applies a bounded packet-local operator before returning the packet to ordinary forwarding. The evaluation records 41,452 decisions, including 17,452 packets processed at multiple nodes, with no incompatible transition or relay image reconstruction. On a fluctuating forest-fire trace, DINA raises deadline tile coverage from 40.4% to 72.6% and receiver accuracy from 77.5% to 95.0% relative to unchanged forwarding. A RescueNet segmentation case raises coverage from 65.6% to 91.7% and dataset-level foreground mIoU from 0.486 to 0.541 on a composite trace. Sufficient-capacity and extreme-capacity profiles locate the measured no-gain and common task-failure boundaries. The contributions are as follows: • We identify and formulate the on-path execution opportunity created when fresh exact-egress state and a still-mutable in-flight packet coexist after that packet has left its source. • We design a self-describing packet and compatible-operator contract that supports independent processing and continued representation evolution across multiple nodes without image reconstruction or cross-packet adaptation state. • We realize the method in an XDP/AF_XDP packet path with an offline-trained, integerscored selector and evaluate its network and receiver-task effects in a 24-node UAV environment across forest-fire classification and RescueNet segmentation workloads.
2
Related Work and Positioning
DINA combines task-aware representation adaptation with execution at the node that is about to transmit the packet. Table 1 compares adjacent approaches by control position, online input, and adaptation object.
3
Table 1: Representative approaches by control position and adaptation object. Line of work
Decision position
Principal online input
Adaptation object
UAV emergency communication [7, 2, 1, 3]
Routing, radio, or endpoints
Topology, propagation, channel, and resource state
Route, coverage, channel use, or endpoint code
ABR, layered coding, and proxy adaptation [8, 9, 10, 11, 12]
Server, client, or edge proxy
Throughput, playback buffer, compute, or edge-link rate
Segment bitrate, coded layer, or removable packet chunk
Task-aware visual analytics [13, 14, 4, 15, 5, 6]
Camera, source, or analytics edge
Content profile, task accuracy, bandwidth, and compute
Frame selection, resolution, model configuration, or encoding allocation
Semantic communication and split inference [16, 17, 18, 19, 20, 21]
Learned endpoints or a DNN partition
Channel state, task loss, and model state
Channel symbols, latent features, or partition point
Programmable Switch, NIC, kernel, in-network or attached worker computing [22, 23, 24, 25, 26, 27, 28]
Application state and data-plane resources
Cached values, aggregates, queries, or packets
Network-side content adaptation [29, 12]
Base station or edge function
Local capacity, queue state, and content priority
Marked message, enhancement layer, or packet chunk
DINA
Capable node before its current egress
Current packet, representation, Current spatial-packet payload and and node-local observation representation type
2.1
UAV Emergency Networks
Flying ad hoc networks and civil UAV systems use airborne nodes as sensors, relays, and temporary infrastructure under mobility and topology change [7, 2]. Emergency-fleet, airborne-coverage, and forest-channel studies further characterize harsh deployment conditions [1, 30, 3]. This literature establishes DINA’s dynamic multi-hop setting. DINA takes the route supplied by the underlying network and acts on the current packet when its next-hop condition changes; its adaptation object is the packet representation rather than the topology, radio resource, or route.
2.2
Endpoint and Task-Aware Adaptation
Adaptive streaming changes data controlled at an endpoint: scalable coding and DASH expose layers or segments, while Pensieve and Swift learn policies or use layered neural codecs [8, 9, 10, 11]. Task-aware visual analytics adapts model configuration, frame selection, resolution, encoding, and region filtering according to resource and receiver utility [13, 14, 4, 15, 5, 6]. These systems establish task-guided reduction at the source, client, or analytics edge. Learned semantic communication instead optimizes transmitted symbols or features for reconstruction or task goals [16, 17, 18, 19], including digital–analog transmission for emergency communication [31]. Collaborative inference partitions a DNN or compresses an intermediate feature between device and edge [20, 21]. DINA lets the receiver task define registered representations, while only a frozen choice and bounded packet transform run on path. The complete task model remains at the receiver, and the local action applies to the current in-flight packet.
2.3
Network-Side Reduction
Edge packet trimming removes less important chunks from layered video using edge capacity [12]. Octopus uses actual 5G base-station capacity to discard application-marked messages or enhance4
ment layers [29]. Both demonstrate that a network-side execution point can react to local service conditions before traffic enters a constrained access link. Recent emergency-network work also places light-model switching inside the network [32], showing a complementary form of task-related local execution. DINA builds on this evidence for network-side actuation but changes a different object. Rather than only dropping a marked unit, removing an enhancement chunk, or changing an inference model, a DINA node converts the payload of the current spatial packet into an application-registered representation. The packet records its resulting type, so a later node can select another compatible transform from the representation actually received. Typed continuation therefore lets one packet’s representation evolve across independently changing hops.
2.4
Programmable In-Network Computing
Programmable data planes execute caching, coordination, telemetry, and distributed-training aggregation in switches or attached processors [22, 23, 24, 25, 26, 27]. P4, XDP, and AF_XDP provide practical substrates for bounded on-path execution [33, 34, 35]; recent systems distribute functions across devices, run learned traffic analysis in a switch, and support payload-mutating functions [28, 36, 37]. DINA supplies an application–network contract for this substrate: a self-describing packet, typed transforms, and a node-local representation policy connect packet processing to a receiver-side machine task.
3
Scenario and Problem Formulation
3.1
Unstable Multi-Hop Image Return
Let the mobile emergency network at time t be a directed graph G(t) = (V, E(t)). A source s ∈ V sends an image object If to destination d ∈ V over the route Pf (t) = (v0 = s, v1 , . . . , vHf = d).
(1)
The underlying network supplies the route and may change it with topology. For directed link e = (v, u), node v reads a local observation ze (t) = ce (t), qe (t), de (t), ℓe (t), . . . , (2) where the components can include current service rate, queued bytes, recent delivery counters, connectivity, or other locally exported measurements. The observation is local to v and is sampled when the packet is processed.
3.2
Where an In-Network Action Can Help
For a packet waiting at node v, changing its representation can affect only links that the packet has not crossed. Let e = (v, u) be its physical next hop. A useful execution opportunity exists when two conditions hold at the same location and time: Fe (t) = I{age(ze (t)) ≤ τe }, M (p, e, t) = I{mutable(p, e, t)}, O(p, v, e, t) = Fe (t)M (p, e, t).
5
(3)
The first term requires a sufficiently fresh application observation. mutable(p, e, t) means that the node can still change the packet before it is committed to egress service. Any on-path node satisfying both terms can execute DINA. This location matters because different hops in a mobile route need not degrade together. Endto-end feedback can describe a path-level consequence, but its observation must return to the source and source-adapted packets must then travel toward the affected hop. If that egress changes within this loop, the source decision may no longer match its condition when those packets arrive. Feedback also cannot modify a packet that has already left the source, and processing after the constrained link cannot reduce the load already offered to it. DINA therefore treats the route as a sequence of local execution opportunities, each defined by the physical egress that the packet is about to traverse. The evaluation measures both network delivery and the receiver task because a smaller payload can still discard useful task evidence.
3.3
Spatial Packet Object
Each 480 × 320 image in the current case is divided into N = 600 16 × 16 spatial packets. Packet pf,i contains pf,i = ⟨f, xi , yi , wi , hi , rf,i , bf,i ⟩, (4) where f is the image identifier, (xi , yi , wi , hi ) is placement geometry, rf,i ∈ R is the current representation, and bf,i is the current payload. Sequence position is not needed for node processing or receiver placement. The packet object separates application knowledge from network execution. The application defines how a spatial region is represented and decoded; the network receives a typed, bounded object that can be transformed without the complete image. Workloads with cross-packet dependencies require an application packet contract that exposes an independently processable object.
3.4
Packet-Local Semantic Adaptation
The application registers a finite set of operators A. Each operator a ∈ A declares an input type, output type, and packet-local transform a : Xrin → Xrout .
(5)
For current representation r, the immutable compatibility registry returns C(r) = {a ∈ A : r ∈ input(a)}.
(6)
When pf,i reaches node v at time tvf,i , DINA chooses v avf,i = πθ pvf,i , z(v,u) (tvf,i ), C(rf,i ) ,
(7)
where u is the packet’s next hop and θ is trained offline and frozen during a run. The packet is then updated as v v bv+ f,i = af,i (bf,i ),
(8)
v+ v rf,i = T (rf,i , avf,i ),
(9)
and forwarded. Retention is an explicit action that leaves the current representation unchanged.
6
The online state used by Eq. (7) ends with the packet. Two packets of the same image can select different operators when they encounter different node-local observations, and coordinatebased placement makes their delivery order irrelevant. For a single packet, the current representation can change at several nodes. Suppose node vj converts RGB24 to FIRE8 and a later node vj+2 receives that packet. The later node filters its choices through C(FIRE8) and may retain FIRE8 or produce a representation whose operator accepts FIRE8. It cannot schedule an RGB-only transform or recreate discarded color channels. Only the current type and payload are needed to determine the next compatible action. If a link recovers, a later RGB24 packet from the same image can remain RGB24 even though an earlier packet was compacted. A subsequent degradation can produce another representation. The realized image therefore follows packet arrivals and local observations without an image-level transition rule.
3.5
Receiver Task and Performance Boundary
At a configured source-relative time Df , the destination initializes a black canvas Ibf (Df ) and overwrites every coordinate for which a packet is available. The representation identifier determines how the payload is decoded into that region. A fixed receiver model h then produces ybf = h(Ibf (Df )).
(10)
Semantic performance is measured against task label yf through the fixed receiver model. The receiver runs the task at the configured evaluation time and records the coordinates and representations present on the canvas when h is invoked. For a fixed image stream, deadline, topology process, and operator library, the operating space contains three empirically distinguishable regions: • a sufficient region in which unchanged RGB forwarding already supports the task; • an intermediate region in which compact packet representations can change delivery and receiver-task outcomes; and • an extreme region in which even the most compact registered representation leaves insufficient task evidence. Their measured boundaries depend on the workload, deadline, queue, registry, and receiver.
4
DINA Design
4.1
Architecture
Fig. 2 shows the execution boundary. The source performs spatial packetization once. Routing forwards each packet through any number of eligible nodes. At a node, XDP identifies DINA traffic and redirects it to the local AF_XDP worker. The worker reads the exact outgoing-link record, selects and executes an operator, and forwards the updated packet. Full machine inference occurs only at the destination.
7
Figure 2: Packet-local representation evolution along a multi-hop path. Each node observes its own outgoing condition, processes only the current packet, updates the packet’s current representation, and returns it to forwarding. No intermediate node reconstructs the image. The ordinary network plane supplies topology, route selection, queueing, and link service. The DINA packet plane performs parsing, compatibility filtering, operator selection, transformation, and representation update. The application supplies operator definitions, offline training material, the coordinate decoder, and the receiver task. A packet can encounter zero, one, or several eligible nodes. Enabled nodes return it to the route supplied by ordinary forwarding after bounded processing. Resolving the next hop before selection binds the worker’s observation to the exact physical egress.
4.2
Self-Describing Packet Contract
Table 2 lists the logical packet fields. Object and geometry fields let the receiver place a packet independently. The current representation makes both downstream interpretation and compatibility checks independent of operation history. Wire length is updated after a transform so the compact representation reduces the load offered to the following link. Let Hp denote the fixed packet and transport overhead and let La (w, h) be the serialized payload produced by representation a for a w × h region. The load offered to the next hop after processing is B(p, a) = Hp + La (w, h). (11) For a set Pe (∆) of packets presented to egress e during interval ∆, the transformed offered load is X Be (∆) = B(p, ap ). (12) p∈Pe (∆)
These equations account for actual serialized bytes; an operator is not credited with a networking benefit merely because its decoded canvas looks sparse. Headers, metadata, and any packet that is retained remain in the wire budget.
4.3
Task-Related Operator Registry
Operators are supplied and profiled by the application. The forest-fire case instantiates the interface with the simple chain in Table 3. A shared color rule marks fire-candidate pixels. FIRE8 retains the 8
Table 2: Logical packet contract. Field
Purpose
Object identifier Coordinates and geometry Current representation Payload length Current payload
Associates independently arriving tiles with one image. Determines placement without arrival-order assumptions. Selects payload decoder and compatible downstream actions. Describes the current wire payload after transformation. Carries the original or task-related spatial representation.
Table 3: Registered representations for a 16 × 16 tile. Rep.
Payload
Accepted inputs
Meaning
RGB24 FIRE8 FIRE1 FIRE1-DS4
768 B 256 B 32 B 2B
RGB24 RGB24, FIRE8 RGB24, FIRE8, FIRE1 all registered types
retain RGB pixels candidate red intensity candidate mask pooled mask
red intensity at those positions, FIRE1 retains the binary candidate mask, and FIRE1-DS4 applies 4 × 4 max pooling to that mask. These operations execute on one tile without neighboring tiles or the receiver model.
Figure 3: Full-scene outputs of the registered forest-fire case-study representations on the same held-out source image. Every panel is produced by applying the current packet-local operator independently to 16 × 16 tiles and reconstructing them by coordinate. From left to right, the four panels show progressively smaller serialized tile payloads; byte counts exclude common headers. Figure 3 makes the application contract concrete. RGB24 preserves the complete scene. FIRE8 preserves one byte of candidate red intensity per pixel. FIRE1 keeps only the candidate predicate, packed as one bit per pixel. FIRE1-DS4 pools each 4 × 4 group inside the tile and therefore carries 16 bits. The receiver expands each representation to its spatial region before running the fixed task model. No intermediate node needs to know whether the scene contains fire; it only executes a registered transform selected by the lightweight policy. Compatibility is encoded as a directed representation graph. A packet can follow RGB24 → FIRE8 → FIRE1 → FIRE1_DS4,
(13)
or skip to any registered output that accepts its current input. Retention at a downstream node is always possible. A transform toward a representation requiring information absent from the current payload is rejected before scoring. 9
Packet-local node procedure Input: packet p, physical egress e, frozen registry and selector. 1. Parse coordinates, current representation r, and payload. 2. Read one consistent current observation ze (t). 3. Construct C(r) from immutable type metadata. 4. Score each a ∈ C(r) using the frozen selector. 5. Select the highest-scoring action; resolve an exact tie in favor of the more informative compatible representation. 6. Apply the action to this payload, update r and wire length, and forward. 7. Discard packet-local variables when the packet leaves the worker.
Figure 4: The DINA online procedure. Each invocation completes for one packet. The graph is immutable during a run and checked independently at every node. The learned policy can retain a packet at one node and compact it at another; a later packet from the same image can make a different choice. A transition is executable when its declared input accepts the current payload.
4.4
Packet Processing Procedure
The critical runtime is summarized in Fig. 4. The deployed selector evaluates frozen learned scores after compatibility filtering. The procedure is deliberately restartable for every packet. All mutable variables—the parsed header view, feature vector, compatible-action mask, scores, and transformed length—belong to the current invocation. Immutable objects such as the operator registry and frozen selector weights can be shared by all packets without creating adaptation history. Queue state belongs to the network observation and may change between invocations; DINA samples it as an input to each packet decision. Representation retention is included in the action set. Exact ties resolve toward the more informative compatible representation, preventing quantized ties from triggering arbitrary compaction. Parser failures, unknown representations, and invalid transforms are counted as implementation errors.
5
Offline Training and Online Selection
DINA uses an offline-trained policy that is frozen for bounded online evaluation. For training item q, ϕq contains packet-visible features and network observations available at an execution node, and Uq,a is the counterfactual utility of compatible action a. Image and trace groups are disjoint across training, validation, and test material. At runtime, a∗ = arg max Qθ a | ϕ(p, ze (t)) . (14) a∈C(r)
The feature function reads the current packet and local observation. The implemented instance uses connectivity, current-to-design rate ratio, exact-egress queue pressure, current-representation rank, and payload ratio. We fit one ridge-regression value function per action and quantize the coefficients into the signed-integer DINA_SELECTOR_V1 ABI. The AF_XDP runtime evaluates D E a∗ = arg max θba , ϕ(p, ze (t)) , (15) a∈C(r)
10
where θba is loaded at worker startup. A parity harness feeds identical current representations and feature vectors to the Python evaluator and compiled C selector; accepted exports match on integer features, compatibility masks, per-action scores, and selected actions. The weights remain immutable throughout each run.
6
Implementation
6.1
Twenty-Four-Node UAV Environment
The maintained environment contains 24 homogeneous network-node containers in a configurable UAV mesh. A topology manager publishes connectivity and directed-link parameters, and the existing routing subsystem derives paths from that topology. Directed-link profiles change independently and asynchronously without resetting queues or application objects. Different nodes on one active route can therefore observe different service regimes, and an in-flight packet can encounter a condition that differs from the one at source emission. The environment records each realized route and node-local worker decision. The primary workload is a continuous stream of independent forest-fire images, with 600 selfcontained packets per image. Congestion changes which spatial packets and representations are available when the receiver constructs each canvas. The RescueNet case retains the same packet geometry and path while using a gray/edge registry and a semantic-segmentation receiver.
6.2
XDP and AF_XDP Packet Path
XDP performs early identification and redirection. An AF_XDP worker inside the network-node container owns the DINA execution step and returns the packet to normal forwarding after processing. AF_XDP provides shared packet rings and user-space access without placing the complete image task in the kernel [35]. The worker contains packet parsers, the compatibility table, four tile operators, and the frozen integer selector. The packet path consists of five bounded stages. XDP classifies an eligible packet and redirects it to the AF_XDP socket. The host forwarding fabric resolves the next hop and inserts a fixed nodelocal selector entry whose argument identifies the physical egress. The worker parses the current representation, reads one consistent observation record, and evaluates the compatible actions. It then rewrites the payload, representation identifier, and length if the chosen action transforms the packet. Finally, the packet is returned to the ordinary forwarding path and the node-local entry is marked as visited for that packet. The visited entry is a packet-local loop guard that prevents immediate redirection to the same worker. If forwarding later reaches another eligible node, that node performs a new decision using its own egress record and the packet’s updated representation. Learned runs load an explicit selector file at startup and record its identity in the run manifest.
6.3
Read-Only Observation Interface
The host exposes a versioned read-only exact-egress table containing connectivity, current and design rate, queue occupancy and capacity, diagnostic counters, sample time, and topology generation. Each worker maps the table once and performs no JSON parsing, socket request, or file write in the packet path. A sequence lock and schema, generation, and next-hop checks prevent workers from consuming a partially updated or mismatched link record. The accepted record identity is retained in the packet event for later audit. 11
6.4
Receiver Reconstruction
The receiver initializes a 480 × 320 black canvas for every object. It decodes each packet according to the packet’s current representation and places the resulting tile at the encoded coordinate, independent of arrival order. At the configured source-relative evaluation time, the primary receiver applies one frozen YOLO model and records the task result. Later arrivals do not rewrite that result. Every network method uses the same receiver model, preprocessing, threshold, coordinate placement, and black-fill rule; intermediate containers neither contain the receiver checkpoint nor query its predictions. The RescueNet receiver uses the same coordinate placement and black-fill rule. Its application registry contains RGB24, Gray8, Gray4, and Edge1, and its frozen DeepLabV3–MobileNetV3 model produces an 11-class segmentation. Training, validation, and test separation for both receivers is specified in Section 7. Packet, forwarding, reconstruction, and task records share one object identifier for end-to-end reconciliation and reproduction of naturally mixed canvases.
7
Evaluation Methodology
The evaluation is organized around the questions in Table 4. Every reported network comparison uses paired inputs and a source-relative decision time. Calibration runs used to locate operating points are excluded from the reported test results. Table 4: Evaluation questions and required evidence. RQ
Question
Evidence
RQ1
Can the fixed receiver interpret registered and naturally mixed representations?
RQ2
Does a packet undergo compatible processing at multiple nodes without image state?
RQ3
Where does network reduction translate into a receiver-task change? Does the mechanism remain responsive under asynchronous link and route conditions?
Held-out task metrics for complete representations, packet mixtures, and black-filled canvases. Per-node transition trace, input/output type audit, local-rate generation audit, and absence of image reconstruction. Paired sufficient, intermediate, and extreme profiles for DINA and unchanged forwarding. Packet representation, queue, coverage, and task outcomes under a continuously varying multi-hop profile. Measured online-decision latency, paired worker CPU and memory, and packet metadata. Frozen RescueNet segmentation model, task-specific registered representations, three static capacity regions, and one composite trace.
RQ4
RQ5
What does packet-local execution cost?
RQ6
Does the same packet path remain useful for a different emergency machine task?
7.1
Dataset and Receiver Model
The forest-fire dataset is divided by source group so that visually derived variants of one source cannot cross training, validation, and held-out test partitions. The receiver is trained on RGB24, FIRE8, FIRE1, FIRE1_DS4, packet-wise mixtures, and black-filled missing regions. Model seed and decision threshold are selected using validation data only, after which the same model and threshold are frozen for every network method.
12
Table 5: Primary forest-fire workload parameters. Parameter
Frozen value
Training/validation/test source images Network-test images (fire/non-fire) Image resolution and tiles Object period / packet pacing Source-relative decision time Queue size and drop policy Topology and routing Link traces Paired test objects per point
2,996 / 736 / 737 40 (20/20) 480 × 320, 600 tiles 13 s / 15 ms 12 s 8 KiB / tail drop 24 UAVs / dynamic two-hop primary routes deterministic, phase-aligned 10-epoch profiles 40
Table 6: Frozen RescueNet network-test contract. Parameter
Frozen value
Official train/validation/test pairs Network-test images Registered representations Receiver metric Image geometry Object period / packet pacing Canvas window / queue Primary / backup route
3,595 / 449 / 450 40 test images RGB24, Gray8, Gray4, Edge1 Foreground mIoU 480 × 320, 600 tiles 13 s / 15 ms 12 s / 8 KiB tail drop 24–4–3–7–17 / 24–14–13–18–17
The bounded second case uses the official RescueNet post-disaster UAV segmentation release [38]. It contains 3,595 training pairs, 449 validation pairs, and 450 reserved test pairs, each with an 11-class mask. Two DeepLabV3–MobileNetV3 candidates receive the same 40-epoch budget on deterministic registered mixtures; validation foreground mIoU selects one checkpoint before any test execution. The formal network playlist contains 40 test images chosen by a content-hash order that does not read labels, model outputs, or network results. Table 6 records the network contract shared by its 24 method–region runs.
7.2
Compared Methods
The primary comparison is between DINA and forwarding-only, which retains each packet’s current representation. Fixed-representation controls traverse the same worker and operator path but apply one named representation without live link features. Every paired run shares topology, route realization, image order, pacing, coordinates, link trace, queue, decision time, canvas construction, and receiver model. Octopus and packet trimming are compared at the mechanism level in Section 2 because their application contracts expose droppable messages or layers instead of transformable spatial packets.
7.3
RQ1: Registered and Mixed Task Inputs
Held-out images are materialized as every complete representation, deterministic packet-wise mixtures, and mixtures with black coordinates. Source-group splits keep all variants of one scene together. We report the receiver confusion matrix, accuracy, precision, recall, false-fire rate, F1, F2, and balanced accuracy; network runs then evaluate their realized canvases directly.
13
7.4
RQ2: Multi-Node Execution Audit
We trace packets processed by at least two nodes and audit the physical egress, topology generation, observation, input/output representation, compatibility mask, and wire length at every decision. A run is valid when transitions are compatible, events reconcile, no relay reconstructs an image, and coordinate placement succeeds under reordered arrival.
7.5
RQ3: Paired Performance Boundary
Validation-only calibration locates sufficient, intermediate, and extreme operating regions. The profiles are then frozen and evaluated on paired held-out images at a source-relative 12-second decision time. Their rates are interpreted together with the configured offered load, packet sizes, queue, and receiver task.
7.6
RQ4: Unstable Links
The intermediate profile changes local links throughout packet emission and can trigger route changes. Each worker decision is joined to its exact-egress rate and queue observation on a common source-relative timeline, together with the packet’s input and selected representation.
7.7
RQ5: Packet-Path Cost
The worker times exact-egress observation, feature construction, selector inference, an optional transform, and checksum update after parsing and before forwarding. Thirty-second container samples provide paired CPU and memory measurements; header sizes come from the compiled packet ABI.
7.8
RQ6: Bounded Cross-Task Reuse
The RescueNet case changes the application registry and receiver task while retaining the packet ABI, coordinate reconstruction, black fill, egress observation, compatibility filtering, selector interface, and XDP/AF_XDP path. DINA, forwarding, and four fixed representations share a 40-image playlist, frozen routes, and sufficient, intermediate, extreme, and composite profiles.
7.9
Metrics and Paired Analysis
Network metrics are tile coverage and payload bytes at the source-relative time, congestion drops, queue occupancy, and token or queue waiting. Task metrics are accuracy, fire precision, fire recall, false-fire rate, F1, balanced accuracy, and the confusion matrix. RescueNet reports dataset-level foreground mIoU from the accumulated pixel confusion matrix and the paired per-image mIoU distribution. For image f , deadline tile coverage is covf (Df ) =
|{i : pf,i is placed by Df }| . N
(16)
Delivered bytes are counted at the receiver by the same time and reported separately from coverage. Paired image identifiers and identical traces support exact sign tests for coverage and McNemar tests for classification correctness. RescueNet foreground mIoU is accumulated over all 40 pixel confusion matrices; paired per-image differences use 10,000 fixed-seed percentile-bootstrap resamples, with undefined pairs reported explicitly. 14
Table 8: Multi-node packet-execution audit. Audit item
Result
Worker packet decisions Packets processed at two or more nodes Actual representation changes Incompatible decisions Exact-egress generation mismatches Relay image-reconstruction events Receiver format or placement errors Unreconciled detour events
41,452 17,452 24,006 0 0 0 0 0
8
Evaluation Results
8.1
Receiver Capability across Registered Inputs
Table 7 evaluates the frozen receiver on 737 source-group-disjoint test images. Each complete representation contributes one canvas per source image. The packet-wise rows aggregate four deterministic mixtures each, either without missing tiles or with 10% and 25% black-filled coordinates. All six input families retain useful fire evidence under the same validation-selected decision threshold. In particular, naturally mixed canvases achieve 95.66% accuracy and 97.21% fire recall; adding the evaluated black-filled masks changes these values only to 95.56% and 96.68%, respectively. These measurements establish the receiver input family used by the case study; each realized network canvas is still evaluated directly. Table 7: Receiver capability on source-group-disjoint test inputs. Input family RGB24 FIRE8 FIRE1 FIRE1-DS4 Packet-wise mixtures Mixtures with black fill
8.2
737 737 737 737 2,948 2,948
0.9213 0.9471 0.9376 0.9362 0.9566 0.9556
Fire recall
F1
False-fire rate
0.8564 0.9909 0.9188 0.9101 0.9789 0.9432 0.8926 0.9789 0.9337 0.8923 0.9758 0.9322 0.9340 0.9721 0.9526 0.9364 0.9668 0.9513
0.1355 0.0788 0.0961 0.0961 0.0560 0.0536
Samples Accuracy Precision
Packet-Local Execution across Multiple Nodes
The execution audit follows 24,000 source packets through a continuously changing 24-node topology. The active primary route alternates between 24 → 18 → 17 and 24 → 23 → 17. Of the source packets, 17,452 are examined by workers at two or more on-path nodes. The trace records 41,452 packetlocal decisions, including 24,006 actual representation changes and 17,446 compatible unchanged outcomes. Later workers consume the representation carried by each packet; no operation-history or image-progress state appears in the worker decision record.
8.3
Paired Network and Task Outcomes
In the sufficient profile, the two-hop service varies between 0.5 and 2.0 Mbit/s. Both methods deliver all 600 coordinates of every image by the 12-second decision time, and neither records a congestion drop. DINA reduces receiver-side payload by 86.18% (2.55 versus 18.43 MB) but cannot increase coverage beyond 100%. Its 92.5% accuracy is close to forwarding-only’s 90.0%; the paired directions 15
are three DINA-only correct and two forwarding-only correct decisions (McNemar p = 1). Thus the sufficient point shows payload reduction without a significant task advantage. The continuously fluctuating intermediate trace repeats 900, 600, 100, 40, 20, 20, 30, 60, 150, and 700 kbit/s epochs. Object release is aligned to the same trace phase in both methods. DINA raises mean deadline coverage from 40.42% to 72.60% while reducing receiver-side payload from 7.45 MB to 0.79 MB. It delivers more coordinates for every one of the 40 paired images, with a mean gain of 193.075 tiles (exact paired sign test, p = 1.82 × 10−12 ). The network change translates into a receiver-task change: accuracy increases from 77.5% to 95.0%, fire recall from 80% to 100%, and false-fire rate decreases from 25% to 10%. DINA alone is correct on seven paired images, while forwarding-only alone is correct on none (exact McNemar p = 0.0156). The extreme profile instead fluctuates between 5 and 37.5 kbit/s. DINA still raises mean deadline coverage from 4.14% to 22.30% and delivers more coordinates for all 40 paired images (mean gain 108.975 tiles, exact sign-test p = 1.82 × 10−12 ). This network gain is not sufficient for reliable fire evidence: fire recall is only 15% with DINA and 30% with forwarding-only, and the paired correctness directions do not establish an advantage (three DINA-only versus six forwarding-only, McNemar p = 0.508). The corresponding 57.5% and 65.0% accuracies are dominated by all 20 non-fire images being classified correctly. This profile therefore lies outside the useful task region despite its coverage gain. Table 9: Main paired network and receiver-task results. Drops are fabric egress-congestion drops over the complete paired run. Region
Method
Sufficient
Forwarding-only DINA Forwarding-only DINA Forwarding-only DINA
Intermediate Extreme
8.4
Coverage
Delivered Fire False-fire bytes Drops Accuracy recall rate
1.0000 18,432,000 0 1.0000 2,547,356 0 0.4042 7,450,368 14,380 0.7260 788,552 6,758 0.0414 763,392 22,522 0.2230 10,706 17,967
0.900 0.925 0.775 0.950 0.650 0.575
1.00 0.95 0.80 1.00 0.30 0.15
0.20 0.10 0.25 0.10 0.00 0.00
Packet Response within a Fluctuating Image Transfer
Figure 5 follows one held-out image through 9.16 seconds of the intermediate profile. The joined trace contains 1,042 decisions at nodes 18, 23, and 24 for 600 source packets. Their observed egress rates fall from 0.6 Mbit/s through 0.1, 0.04, and 0.02 Mbit/s before recovering to 0.9 Mbit/s. Low service and high queue pressure produce mostly FIRE1-DS4 decisions; recovery produces later FIRE1 and FIRE8 decisions. The complete object trace contains 766 FIRE1-DS4, 184 FIRE1, and 92 FIRE8 node decisions.
16
Node 18
Local rate (Mbit/s)
0.8
Node 23
Node 24
0.6 0.4 0.2
Decision share / queue pressure
0.0 1.0
FIRE8
FIRE1
FIRE1-DS4
Queue pressure P95
0.8 0.6 0.4 0.2 0.0
0
2
4 6 Time from source object start (s)
8
Figure 5: Packet-local response during one image transfer in the fluctuating intermediate profile. The upper panel joins each on-path decision to its exact-egress rate. The lower panel bins selected representations and P95 queue pressure every 0.5 seconds.
8.5
Cross-Task Reuse on RescueNet
The RescueNet matrix changes both the task-related representations and the receiver from binary fire classification to 11-class post-disaster semantic segmentation. Each of the 24 method–region runs uses the same 40 test images, 600 source tiles per image, phase alignment, queue, sourcerelative deadline, network image, and receiver checkpoint. The generated primary route is 24 → 4 → 3 → 7 → 17; the installed backup is 24 → 14 → 13 → 18 → 17. No route replacement occurs during a run, and every per-run and cross-run contract check passes. Table 10 gives the DINA and forwarding results plus the fixed representation with the highest observed dataset-level mIoU in each region; all four fixed controls remain in the underlying comparison matrix. At the sufficient point, DINA retains RGB24 and exactly matches forwarding in coverage and mIoU. Gray8, Gray4, and Edge1 all deliver every tile but reduce mIoU, so unnecessary compaction is directly visible.
17
Table 10: Frozen RescueNet network results. The fixed column reports the highest-mIoU fixed representation in each region. DINA Region Sufficient Intermediate Composite Extreme
Forward only
Coverage mIoU Coverage mIoU Representation 1.0000 0.5391 0.7938 0.4846 0.9165 0.5405 0.0415 0.0080
1.0000 0.5391 RGB24 0.3110 0.3738 Gray4 0.6561 0.4859 Gray8 0.0291 0.0078 Gray8
Highest-mIoU fixed Coverage
mIoU
1.0000 0.8158 0.8170 0.0373
0.5391 0.5053 0.5312 0.0077
The static intermediate point isolates a strong in-network processing gain over unchanged forwarding. DINA raises mean deadline coverage by 0.4827 with a paired 95% interval of [0.4651, 0.5007]. Across the 39 images with a defined pair, the mean per-image foreground-mIoU difference is 0.1605 with interval [0.0978, 0.2315]; DINA is higher on 35 and lower on four. Fixed Gray4 nevertheless exceeds DINA at this one static point in both aggregate coverage and aggregate mIoU. The composite trace below evaluates the selector’s intended role under changing conditions. The composite trace tests the reason for online selection. DINA produces 36 naturally mixed canvases among 40 images and raises coverage from 0.6561 to 0.9165 relative to forwarding. Its dataset-level foreground mIoU is 0.5405, compared with 0.4859 for forwarding. The paired mean differences are 0.2604 in coverage, with interval [0.2117, 0.3095], and 0.1089 in per-image mIoU, with interval [0.0547, 0.1747]. DINA also has a higher aggregate mIoU than every fixed representation. Against Gray8, the highest-mIoU fixed control, the paired per-image mean difference is 0.0166 with interval [−0.0043, 0.0374], leaving the task advantage over this fixed control statistically inconclusive. Edge1 reaches 0.9985 coverage but only 0.3542 mIoU, demonstrating that minimizing bytes or maximizing tile count alone is not the task objective. Figure 6 shows how the composite aggregate arises. The received DINA tiles alternate between retained RGB24 and compact Gray4 as the phase-aligned service process changes; Gray8 appears only briefly. The lower panel pairs each image with forwarding under the identical release phase. Decisions occur at four capable nodes on the primary path, and later nodes consume only the current representation carried by each packet. The response is consequently the realized sequence of independent node-local choices.
18
Figure 6: Per-image response in the RescueNet composite trace. The upper panel shows the representation shares among DINA tiles received by the deadline; the lower panel shows paired deadline coverage for DINA and forwarding only. Each point uses the same frozen image order and trace phase. The PNG is generated by Python from per-image formal evidence. At the extreme point, every method yields a dataset-level foreground mIoU between 0.0064 and 0.0080. Edge1 raises coverage only to 0.0456, and no paired task comparison establishes a useful advantage. This defines the common physical and semantic failure boundary for the frozen packet rate, deadline, queue, registry, and receiver.
8.6
Online Processing Cost
Table 11 summarizes all 41,452 worker decisions in the continuous trace. The median measured decision path is 3.129 µs and the P95 is 6.935 µs. Compatible transformations cost more than an unchanged decision, with a P95 of 7.843 µs across the three output formats. The 80.23 decisions/s value is the offered rate of this experiment, not a worker saturation limit. Across the same 30-second resource samples, the three eligible node containers use 1.86 CPU percentage points and 7.3 MiB more on average than forwarding-only. The packet carries a fixed 64-byte SPP header, including a one-byte current representation identifier and no operation history.
9
Discussion
9.1
When Node-Local Adaptation Helps
DINA uses the co-location expressed by Eq. (3): a node has an actionable observation of its egress while the affected packet remains mutable before that egress. In the target high-dynamic regime, an endpoint observe–return–act loop cannot reliably track independent intermediate-egress variation:
19
Table 11: Measured packet-path cost under the continuous workload. Metric
Measured value
All decisions, median / P95 Unchanged, median / P95 Transformed, median / P95 FIRE8/FIRE1/FIRE1-DS4 transform P95 Observed decisions Eligible-node mean CPU, DINA / forward Eligible-node mean memory, DINA / forward SPP header / current representation field Operation history / relay image buffer
3.129 / 6.935 µs 1.355 / 2.616 µs 3.982 / 7.843 µs 5.431 / 5.968 / 8.340 µs 41,452 (80.23 s−1 ) 6.20% / 4.34% 110.9 / 103.6 MiB 64 B / 1 B 0B/0B
the reported condition may change before source-adapted packets reach that hop. The on-path node instead samples the current local condition and acts on the packet already present there. Source adaptation remains effective when the relevant state is stable and the source still owns the affected data. Once a packet reaches a changing intermediate hop, node-local transformation remains actionable before egress service; a transform after the weak hop cannot undo prior queueing or loss. Across multiple capable nodes, the current representation lets each node continue compatible processing without coordinating an image action.
9.2
Operating and Semantic Boundaries
The measured regions expose the method’s operating boundary. With sufficient capacity, unchanged forwarding already delivers the task evidence. In the intermediate region, compact representations change delivery and the receiver task. Under extreme capacity, even the smallest registered representation leaves inadequate spatial evidence. Semantic correctness is the outcome of the fixed receiver task on the realized canvas, reported together with the network delivery that produced it. The forest-fire color operators and RescueNet gray/edge operators retain different task evidence. Both use the same typed packet, local observation, compatibility filtering, and on-path execution, while each application supplies its registry, training material, and receiver. The RescueNet static intermediate result also shows that a fixed compact representation can exceed the learned selector at one stable operating point; the selector’s measured role is adaptation across changing conditions.
9.3
Scope and Deployment
The continuous independent-image stream exercises offered load, queueing, loss, route and link transitions, and mixed receiver canvases without adding inter-frame codec dependencies. Video deployment would require a codec-aware packet contract exposing independently processable objects. Other machine tasks similarly provide their own packetization, registry, receiver, and held-out task validation. Payload transformation operates within a trusted domain shared by the application and network. The host publishes link state through a read-only table, the application authorizes a finite packet format and operator set, and opaque encrypted payloads require an explicitly transformable region. The controlled 24-node environment tests dynamic links, finite queues, routes, and containerized workers reproducibly; field deployment additionally requires radio-state, energy, security, and operational validation.
20
10
Conclusion
DINA turns an on-path node’s local observation into an immediate action on the packet already present at that node. Self-describing spatial packets and typed operator compatibility allow each packet to be processed independently and to continue evolving across a multi-hop route. The online node limits its work to a frozen lightweight selector and a registered packet transform; full task inference remains at the destination. The destination measures semantic success through its fixed machine task on the coordinate-reconstructed canvas. This design provides a concrete innetwork-computing mechanism for studying where local representation changes expand the usable operating region of unstable mobile emergency networks. A forest-fire classifier and a RescueNet post-disaster segmenter show this packet-path capability under two task-specific registries, while sufficient and extreme profiles delimit where local adaptation provides no additional task benefit or cannot overcome the physical channel.
Author Contributions Conceptualization, Z.R. and W.C.; methodology, Z.R.; software, Z.R.; validation, Z.R. and T.Z.; formal analysis, Z.R. and T.Z.; investigation, Z.R.; data curation, Z.R.; visualization, Z.R.; writing— original draft preparation, Z.R.; writing—review and editing, Z.R., T.Z. and W.C.; supervision, W.C.; project administration, W.C. All authors have read and approved this manuscript.
Funding This research was funded by the National Key Research and Development Program of China, grant number 2023YFC3011502.
Institutional Review Board Statement Not applicable.
Informed Consent Statement Not applicable.
Data Availability Statement The data and code supporting the findings of this study are available from the corresponding author upon reasonable request.
Acknowledgments The background illustration in Figure 1 was generated using OpenAI’s image-generation system from an author-directed prompt.
21
Conflicts of Interest The authors declare no conflicts of interest. The funder had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.
References [1] Zhuohui Yao, Wenchi Cheng, Wei Zhang, Tao Zhang, and Hailin Zhang. The rise of UAV fleet technologies for emergency wireless communications in harsh environments. IEEE Network, 36(4):28–37, 2022. [2] Samira Hayat, Evsen Yanmaz, and Raheeb Muzaffar. Survey on unmanned aerial vehicle networks for civil applications: A communications viewpoint. IEEE Communications Surveys & Tutorials, 18(4):2624–2661, 2016. [3] Rong Yuan, Wenchi Cheng, Wei Guo, Dan Fei, Haoran Chen, Yayun Ao, and Yudong Fang. Channel measurement and modeling for 380 MHz V2I emergency communications in forested scenarios. Journal of Communications and Information Networks, 10(3):276–286, 2025. [4] Ben Zhang, Xin Jin, Sylvia Ratnasamy, John Wawrzynek, and Edward A. Lee. AWStream: Adaptive wide-area streaming analytics. In 2018 Conference of the ACM Special Interest Group on Data Communication, pages 236–252, 2018. [5] Kuntai Du, Qizheng Zhang, Anton Arapin, Haodong Wang, Zhengxu Xia, and Junchen Jiang. AccMPEG: Optimizing video encoding for accurate video analytics. Proceedings of Machine Learning and Systems, 4:450–466, 2022. [6] Tao Chen, Yanling Bu, Yue Zeng, Lei Xie, and Sanglu Lu. RegionFilter: Region-aware video filtering mechanism on resource-constrained edge nodes. Computer Networks, 251:110624, 2024. [7] Ilker Bekmezci, Ozgur Koray Sahingoz, and Samil Temel. Flying ad-hoc networks (FANETs): A survey. Ad Hoc Networks, 11(3):1254–1270, 2013. [8] Heiko Schwarz, Detlev Marpe, and Thomas Wiegand. Overview of the scalable video coding extension of the H.264/AVC standard. IEEE Transactions on Circuits and Systems for Video Technology, 17(9):1103–1120, 2007. [9] Thomas Stockhammer. Dynamic adaptive streaming over HTTP: Standards and design principles. In Second Annual ACM Conference on Multimedia Systems, pages 133–144, 2011. [10] Hongzi Mao, Ravi Netravali, and Mohammad Alizadeh. Neural adaptive video streaming with Pensieve. In 2017 ACM SIGCOMM Conference, pages 197–210, 2017. [11] Mallesham Dasari, Kumara Kahatapitiya, Samir R. Das, Aruna Balasubramanian, and Dimitris Samaras. Swift: Adaptive video streaming with layered neural codecs. In 19th USENIX Symposium on Networked Systems Design and Implementation, pages 103–118, 2022. [12] Mustafa Tüker, Emre Karakış, Müge Sayıt, and Stuart Clayman. Using packet trimming at the edge for in-network video quality adaption. Annals of Telecommunications, 79(3–4):197–210, 2024. 22
[13] Haoyu Zhang, Ganesh Ananthanarayanan, Peter Bodik, Matthai Philipose, Paramvir Bahl, and Michael J. Freedman. Live video analytics at scale with approximation and delay-tolerance. In 14th USENIX Symposium on Networked Systems Design and Implementation, pages 377–392, 2017. [14] Junchen Jiang, Ganesh Ananthanarayanan, Peter Bodik, Siddhartha Sen, and Ion Stoica. Chameleon: Scalable adaptation of video analytics. In 2018 Conference of the ACM Special Interest Group on Data Communication, pages 253–266, 2018. [15] Yuanqi Li, Arthi Padmanabhan, Pengzhan Zhao, Yufei Wang, Guoqing Harry Xu, and Ravi Netravali. Reducto: On-camera filtering for resource-efficient real-time video analytics. In 2020 Annual Conference of the ACM Special Interest Group on Data Communication, pages 359–376, 2020. [16] Eirina Bourtsoulatze, David Burth Kurka, and Deniz Gündüz. Deep joint source-channel coding for wireless image transmission. IEEE Transactions on Cognitive Communications and Networking, 5(3):567–579, 2019. [17] Huiqiang Xie, Zhijin Qin, Geoffrey Ye Li, and Biing-Hwang Juang. Deep learning enabled semantic communication systems. IEEE Transactions on Signal Processing, 69:2663–2675, 2021. [18] Emilio Calvanese Strinati and Sergio Barbarossa. 6G networks: Beyond shannon towards semantic and goal-oriented communications. Computer Networks, 190:107930, 2021. [19] Elif Uysal, Onur Kaya, Anthony Ephremides, James Gross, Marian Codreanu, Petar Popovski, Mohamad Assaad, Gianluigi Liva, Andrea Munari, Beatriz Soret, Touraj Soleymani, and Karl Henrik Johansson. Semantic communications in networked systems: A data significance perspective. IEEE Network, 36(4):233–240, 2022. [20] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. ACM SIGARCH Computer Architecture News, 45(1):615–629, 2017. ASPLOS 2017. [21] Jiawei Shao and Jun Zhang. BottleNet++: An end-to-end approach for feature compression in device–edge co-inference systems. In 2020 IEEE International Conference on Communications Workshops, pages 1–6, 2020. [22] Amedeo Sapio, Ibrahim Abdelaziz, Abdulla Aldilaijan, Marco Canini, and Panos Kalnis. Innetwork computation is a dumb idea whose time has come. In 16th ACM Workshop on Hot Topics in Networks, pages 150–156, 2017. [23] Ming Liu, Liang Luo, Jacob Nelson, Luis Ceze, Arvind Krishnamurthy, and Kishore Atreya. IncBricks: Toward in-network computation with an in-network cache. In Twenty-Second International Conference on Architectural Support for Programming Languages and Operating Systems, pages 795–809, 2017. [24] Xin Jin, Xiaozhou Li, Haoyu Zhang, Robert Soule, Jeongkeun Lee, Nate Foster, Changhoon Kim, and Ion Stoica. NetCache: Balancing key-value stores with fast in-network caching. In 26th Symposium on Operating Systems Principles, pages 121–136, 2017.
23
[25] Xin Jin, Xiaozhou Li, Haoyu Zhang, Nate Foster, Jeongkeun Lee, Robert Soule, Changhoon Kim, and Ion Stoica. NetChain: Scale-free sub-RTT coordination. In 15th USENIX Symposium on Networked Systems Design and Implementation, pages 35–49, 2018. [26] Arpit Gupta, Rob Harrison, Marco Canini, Nick Feamster, Jennifer Rexford, and Walter Willinger. Sonata: Query-driven streaming network telemetry. In 2018 Conference of the ACM Special Interest Group on Data Communication, pages 357–371, 2018. [27] Amedeo Sapio, Marco Canini, Chen-Yu Ho, Jacob Nelson, Panos Kalnis, Changhoon Kim, Arvind Krishnamurthy, Masoud Moshref, Dan R. K. Ports, and Peter Richtarik. Scaling distributed machine learning with in-network aggregation. In 18th USENIX Symposium on Networked Systems Design and Implementation, pages 785–808, 2021. [28] Changgang Zheng, Haoyue Tang, Mingyuan Zang, Xinpeng Hong, Aosong Feng, Leandros Tassiulas, and Noa Zilberman. DINC: Toward distributed in-network computing. Proceedings of the ACM on Networking, 1(CoNEXT3):1–25, 2023. Article 14. [29] Yongzhou Chen, Ammar Tahir, Francis Y. Yan, and Radhika Mittal. Octopus: In-network content adaptation to control congestion on 5G links. In 8th ACM/IEEE Symposium on Edge Computing, pages 199–214, 2023. [30] Akram Al-Hourani, Sithamparanathan Kandeepan, and Simon Lardner. Optimal LAP altitude for maximum coverage. IEEE Wireless Communications Letters, 3(6):569–572, 2014. [31] Yuzhou Fu, Wenchi Cheng, Jingqing Wang, Liuguo Yin, and Wei Zhang. Digital-analog transmission based emergency semantic communications. arXiv preprint arXiv:2501.01616, 2025. [32] Yuehan Li, Zhiyuan Ren, Tao Zhang, and Wenchi Cheng. In-network artificial computing enhanced light model-switching for emergency communications networks. arXiv preprint arXiv:2605.10070, 2026. [33] Pat Bosshart, Dan Daly, Glen Gibb, Martin Izzard, Nick McKeown, Jennifer Rexford, Cole Schlesinger, Dan Talayco, Amin Vahdat, George Varghese, and David Walker. P4: Programming protocol-independent packet processors. ACM SIGCOMM Computer Communication Review, 44(3):87–95, 2014. [34] Toke Høiland-Jørgensen, Jesper Dangaard Brouer, Daniel Borkmann, John Fastabend, Tom Herbert, David Ahern, and David Miller. The eXpress Data Path: Fast programmable packet processing in the operating system kernel. In 14th International Conference on Emerging Networking Experiments and Technologies, pages 54–66, 2018. [35] The Linux Kernel Developers. AF_XDP — the linux kernel documentation. https://docs. kernel.org/networking/af_xdp.html, 2026. Accessed: 2026-08-31. [36] Jinzhu Yan, Haotian Xu, Zhuotao Liu, Qi Li, Ke Xu, Mingwei Xu, and Jianping Wu. Brainon-Switch: Towards advanced intelligent network data plane via NN-Driven traffic analysis at Line-Speed. In 21st USENIX Symposium on Networked Systems Design and Implementation, pages 419–440, 2024. [37] Tao Ji, Rohan Vardekar, Balajee Vamanan, Brent E. Stephens, and Aditya Akella. MTP: Transport for In-Network computing. In 22nd USENIX Symposium on Networked Systems Design and Implementation, pages 959–977, 2025. 24
[38] Maryam Rahnemoonfar, Tashnim Chowdhury, and Robin Murphy. RescueNet: A high resolution UAV semantic segmentation dataset for natural disaster damage assessment. Scientific Data, 10(1):913, 2023.
25