Semantic Split Inference for Remote Modulation Recognition

1

Semantic Split Inference for Remote Modulation Recognition

arXiv:2610.07682v1 [cs.NI] 6 Oct 2026

Samer Lahoud, Senior Member, IEEE and Kinda Khawam

Abstract—Remote automatic modulation recognition balances sensing-node complexity, reporting cost and accuracy. To address this trade-off, we propose channel-aware semantic split inference: a sensing node sends a semantic report over a noisy link and the edge server completes recognition. In our model, split depth and report length are independent design variables, with endto-end training through the channel. We compare the resulting design against basic split placements, which run inference at the edge server or at the sensing node, and against a state-ofthe-art collaborative scheme. We assess sensing-node model size, computation, latency and energy against recognition accuracy. We show that intermediate splits give the best accuracy–cost trade-off. Index Terms—Semantic communication, split inference, automatic modulation recognition, task-oriented communication.

I. I NTRODUCTION Future wireless networks will require continuous awareness of increasingly congested and heterogeneous spectrum. Automatic modulation recognition (AMR) identifies the modulation format of an observed waveform and supports dynamic spectrum access, interference analysis, and anomalous-signal detection [1], [2]. In large-scale monitoring systems, resourceconstrained sensing nodes observe the spectrum and an edge server performs recognition. Forwarding raw in-phase and quadrature (I/Q) samples preserves the sensed waveform but requires substantial uplink resources. Running the complete classifier locally and reporting its output reduces this load but requires the full model at every node. This letter introduces channel-aware semantic split inference, a task-oriented communication framework for remote AMR [3]. The recognition network is a neural classifier that maps a frame of I/Q samples to a modulation label. We partition it between the sensing node and the edge server at the split depth, the network stage at which the report is formed. There, the node maps its intermediate features to a compact semantic report and sends it over a noisy reporting channel; the edge server completes inference [4]. The report is optimized for classification rather than waveform reconstruction. This yields two coupled design questions: where in the network should the report be formed, and how long should it be? A shallow split limits the sensing-node model but produces a weakly processed representation, whereas a deep split produces a more task-specific representation at greater local cost. Report length sets the number of channel uses and affects recognition accuracy. Samer Lahoud is with the Faculty of Computer Science, Dalhousie University, Halifax, Canada (email: sml@dal.ca). Kinda Khawam is with LISN, Université Paris-Saclay, Gif-sur-Yvette, France (email: kinda.khawam@lisn.fr)

Two limiting cases bound this choice. In one, the edge server performs almost all of the inference and the node forwards a barely processed representation. In the other, the node performs almost all of it and reports only a decision. State-of-the-art collaborative-AMR schemes operate between these limits but fix where the report is formed, so that report length has no independent effect [5]. The contributions of this letter are threefold. First, we formulate remote AMR with distinct sensing and reporting channels and evaluate both AWGN and flat Rayleigh reporting links. Second, we vary split depth and report length independently, including the allocation of a fixed length between semantic width and temporal resolution. We characterize node model size, computation, measured latency and energy, and accuracy. Third, under matched conditions we compare the selected report against both basic split placements and a stateof-the-art collaborative scheme. We show that intermediate splits provide the best accuracy–cost trade-off at a matched reporting rate. II. R ELATED W ORK AND P OSITIONING Deep learning has become the dominant approach to AMR: classifiers trained directly on raw I/Q samples learn discriminative temporal representations and outperform hand-designed feature pipelines [2]. The architectures and signal representations that make this work are extensively surveyed [1], [6]. Attention has also shifted from accuracy alone to the cost of obtaining it, as in recognizers co-designed with hardware accelerators for constrained platforms [7]. This literature nevertheless assumes that the observation and the classifier sit on the same device. That assumption fails in large-scale monitoring, where the design problem extends from the classifier to the representation each node transmits. Federated learning keeps observations local and exchanges model parameters instead [8]. This addresses privacy and training cost, but still requires the complete model at every node at inference time. Split computing takes the complementary route of partitioning a trained network between device and server, so that the position of the cut and the size of the transmitted tensor become the governing design choices [9]. In this setting, the link is usually treated as an error-free bit pipe. Two AMR schemes follow this route and are closest to ours. C-AMC places a compression network at the sensing node and a recurrent classifier at the edge server [5]. Its node network pools features to a fixed width before the transmit projection, so the nominal report length can exceed the intrinsic dimension of the representation it carries. SplitAMC [10] instead splits a residual backbone

2

after its first or second stage and forwards the intermediate feature in full. This exposes no report-length variable and requires at least eight times as many channel uses as raw I/Q forwarding. SplitAMC is also the uncompressed baseline against which [5] reports. In neither case are split depth and report length independent design variables, so the accuracy obtained at a given reporting rate cannot be attributed to the split placement alone. Separating the two is the starting point of this letter. Our approach draws on task-oriented communication, which optimizes the transmitted representation for a downstream task rather than for exact source reconstruction [3], [11]. Split inference realizes this directly for classification: partitioning the recognition network lets the node transmit a representation already shaped for the decision, with both sides trained through a stochastic link [4], [12]. The length of that representation can itself be learned, for instance through an informationbottleneck penalty that prunes uninformative dimensions [13]. We instead fix it by construction, which allows it to be swept against split depth. Because the report is sent as analog amplitudes rather than as a decoded bit stream, its accuracy degrades gradually as the reporting channel worsens. Existing split-inference systems, however, classify clean images [4], [12], [13]. In remote AMR, the node input is itself a noisy observation of an unknown waveform, so accuracy depends on both a sensing and a reporting channel. The report can also retain temporal structure, so its length can be divided between semantic width and token count. III. S YSTEM M ODEL AND S EMANTIC S PLIT I NFERENCE a) Sensing and reporting model: A sensing node observes a complex-baseband frame of N I/Q samples from a signal with unknown modulation class y ∈ {1, . . . , C}. The discrete-time observation is modeled as r[n] = sy [n] ej(2π∆f nTs +Θ) + ηs [n],

n = 1, . . . , N, (1)

where sy [n] is the waveform sample associated with class y, ∆f and Θ are residual carrier-frequency and phase offsets, Ts is the sampling period, and ηs [n] ∼ CN (0, σs2 ) is complex additive white Gaussian noise. The sensing SNR γs = 10 log10 (Ps /σs2 ), with Ps the average signal power, characterizes the quality of the observed waveform. How each impairment of (1) is realized is stated with (5). Stacking the inphase and quadrature components of r = [r[1], . . . , r[N ]] gives the real-valued network input x = [ℜ{r}; ℑ{r}] ∈ R2×N , so that raw forwarding requires Lx = 2N real channel uses. The overall pipeline is shown in Fig. 1. The recognition network contains K residual stages followed by a classification head. A split at depth k ∈ {1, . . . , K} assigns the first k stages to the sensing-node encoder Ek and the remainder to the edge server. A transmit interface Qk maps the encoder output uk = Ek (x) to an unnormalized report bk = Qk (uk ) ∈ RLz , which is normalized independently for each frame: p bk ∥zk ∥22 , = 1. (2) zk = Lz ∥bk ∥2 Lz The semantic report zk thus has length Lz real channel uses at unit average power per use. As in analog joint source– channel coding [14], the normalization prevents unconstrained

AWGN channel, γr η r ∼ N (0, σr2 I)

sensing node γs

x

Encoder Ek

uk TX Qk + norm zk

stages 1:k

nc ×T

+ e zk

yb

Decoder Gk

bk u

stages k+1:K, head

RX Rk

edge server

Fig. 1. Semantic split inference for remote AMR. The sensing node sends a power-normalized report of Lz = nc T real channel uses over a noisy channel (AWGN shown); the edge server completes inference.

amplitude scaling from artificially inflating the effective reporting SNR. One real channel use is one real dimension of the transmitted baseband waveform, so the report occupies Lz /2 complex symbols. The report is sent as uncoded analog amplitudes, and reporting cost is therefore measured in channel uses throughout. It is transmitted over an AWGN channel:
e zk = zk + η r , η r ∼ N 0, σr2 ILz , σr2 = 10−γr /10 , (3) where γr denotes the reporting SNR in decibels. The link is also evaluated under flat Rayleigh block fading with perfect channel state information at the edge server, specified in Sec. IV. Path loss, shadowing and co-channel interference are not modeled separately. For a fixed realization, each only shifts the average received power or the effective noise floor, and is therefore absorbed into γr . For interference, this assumes that the interferer is independent of the report and that the edge server performs no interference cancellation. The reporting SNR γr then reads as a signal-to-interference-plus-noise ratio. At the edge server, the receive interface Rk maps the bk = equalized report b zk to the reconstructed feature tensor u Rk (b zk ), which has the dimensions of uk . On the AWGN channel, b zk = e zk . The remaining residual stages and the classification head Gk then give pk = softmax(Gk (b uk )) and yb = arg maxc pk,c . When k = K, all residual stages are located at the sensing node and Gk contains only the classification head. The split depth k sets the model size and computation assigned to the sensing node, whereas Lz sets the reporting cost. Relative to raw forwarding, the compression ratio is ρ = Lx /Lz = 2N/Lz . The design (k, nc , T ) therefore trades sensing-node model size and computation against reporting cost and recognition accuracy. b) Backbone and semantic bottleneck: We adopt a onedimensional residual CNN based on the raw-I/Q AMR architecture in [2]. The feature extractor contains K residual stages of width w = 64. Each stage applies a 1 × 1 channel projection, two residual convolutional units, and factor-two max pooling, so the encoder output at split k is uk ∈ Rw×Lk with Lk = N/2k . Global average pooling and a two-layer classification head produce the C output logits. The transmit interface applies a 1×1 projection reducing the feature width from w to nc , and an adaptive average pooling PT reducing the temporal length from Lk to T :

bk = vec PT Convtx , Lz = nc T, (4) 1×1 (uk )

3

where φ ∼ U[0, 2π) and τ ∼ U{0, . . . , N − 1}; both transformations preserve the modulation class and frame power. Each impairment of (1) is exercised exactly once. The data generator realizes the sensing noise ηs and the carrier-frequency offset ∆f , drawing an independent center frequency for every frame (Sec. IV). The residual phase is supplied by the rotation, so that Θ = φ. The circular shift Sτ has no counterpart in (1): it represents the unknown position of the signal within the observation window. IV. R ESULTS We evaluate the proposed framework on a balanced, 16class narrowband dataset generated with TorchSig [15]. The dataset spans ASK, PSK, QAM, FSK, OFDM, and analog modulation families. Each frame contains N = 1024 complexbaseband samples normalized to unit average power. During generation, additive noise is applied at sensing SNR γs , while the center frequency and bandwidth are randomized within the observation band. At k = 3, we sweep the full grid nc ∈ {2, 4, 8, 16}, T ∈ {1, 2, 4, 8}, which gives report lengths Lz = nc T from 2 to 128 real channel uses. Accuracy does not increase beyond Lz = 64, so results are shown for Lz ≤ 64, that is, ρ ≥ 32. The split depth k ∈ {1, . . . , 5} is swept at (nc , T ) ∈ {(4, 4), (8, 4)}. Except in the last study of this section, all models are trained over γs ∈ [15, 20] dB and evaluated at γs = 20 dB. In this regime the observation supports reliable recognition, so accuracy differences between configurations reflect the split and the reporting link. The

reporting range γr ∈ [0, 20] dB extends from a link whose noise power equals the report power to one that no longer limits accuracy. Unless stated otherwise, the reporting SNR is sampled as γr ∼ U[0, 20] dB during training. All learned schemes use Adam with an initial learning rate of 10−3 reduced on plateau, a batch size of 256, and a 100-epoch training budget. The class-balanced training, validation, and test sets contain 24 000, 4 800, and 6 400 frames, respectively. Their partitions are fixed across runs. Accuracy results are means over ten independent training runs. Where an uncertainty is given, it is the sample standard deviation (SD) across those runs, in percentage points when shown in parentheses. The complete local classifier has 0.28 M parameters, requires 203 MFLOPs per frame, and achieves a mean accuracy of 0.92 (SD 0.01) at γs = 20 dB. It serves as the no-reportingchannel reference. 1

mean recognition accuracy

after which (2) produces the transmitted report zk . The parameters nc and T , which we call the semantic width and the token count, allocate the report length: T = 1 produces a single global representation, whereas T > 1 preserves a sequence of temporal feature tokens. At the edge server, Rk reshapes b zk into an nc ×T tensor, a learned transposed convolution restores the temporal length Lk , and a 1 × 1 receive projection restores b k . A separate model is initialized the feature width w, giving u and trained for each evaluated (k, nc , T ) configuration. c) Basic split placements: The two extreme split depths define the basic comparisons used in the following. At k = 1, the node model is as small as the parameterization allows, but the report is formed from a barely processed representation. At k = K, the entire backbone runs at the node, and no residual stage is left to process the report after the channel. d) Training: The trainable modules Ek , Qk , Rk , and Gk are optimized end to end through the stochastic reportingchannel layer by minimizing the cross-entropy LCE = E(x,y), ηr [− log pk,y ]. For SNR-aware training, the reporting SNR is redrawn for every mini-batch as γr ∼ U [γmin , γmax ], with an independent noise realization applied to each frame. The channel of (3) enters training as a parameter-free stochastic layer. Its Jacobian with respect to zk is the identity, so gradients reach Qk and Ek unchanged and no surrogate or straight-through estimator is required. We further apply random phase rotation and circular time shifting:   cos φ − sin φ ′ x = R(φ)Sτ (x), R(φ) = , (5) sin φ cos φ

2

split depth k 3

4

5

ρ = 64 ρ = 128

0.90 0.85 0.80 0.75

2

3

4

5

6

7

8

9

measured node inference time [ms] Fig. 2. Mean recognition accuracy versus measured sensing-node inference time (Table I) for split depths k = 1, . . . , 5 and compression ratios ρ ∈ {64, 128} (γs = γr = 20 dB). Error bars are ±1 sample SD.

a) Effect of split depth: Fig. 2 compares mean recognition accuracy across the K = 5 split depths for ρ ∈ {64, 128}. For both ratios, accuracy rises steeply from k = 1 to k = 3, by about 12.5 percentage points. This suggests that the shallow encoder does not form a sufficiently discriminative representation before compression. The subsequent gain from k = 3 to k = 4 is 0.7 points at ρ = 128, within one SD, and 1.6 points at ρ = 64. The lower-compression configuration ρ = 64 is more accurate at every split. Both follow the same increase–plateau– decrease pattern, identifying k ∈ {3, 4} as the strongest region. From k = 4 to k = 5, accuracy decreases by 6 percentage points at both ratios, as no residual stage remains at the edge server to process the noisy report. Recognition therefore depends on both the representation formed before transmission and the edge-side capacity after it. FLOP counts alone do not predict the cost of a split on a deployed node, since stages of comparable arithmetic cost differ in memory traffic and in parallelism. We therefore measured the per-frame sensing-node cost of each split on an NVIDIA Jetson AGX Orin at 30 W, reported in Table I. At k = 3, the node accounts for 90% of the FLOPs but only 58% of the inference time. Energy follows the measured time rather than the FLOP count: the node consumes 6.8 mJ at that split, about 61% of the 11.1 mJ of the complete local classifier.

4

TABLE I S ENSING - NODE COST PER FRAME ON AN NVIDIA J ETSON AGX O RIN (30 W, FP32, BATCH SIZE ONE ): MEDIAN TIME OVER 1000 INFERENCES WITH INTERQUARTILE RANGE (IQR), AND DYNAMIC ENERGY ABOVE IDLE . P ERCENTAGES ARE OF THE LOCAL CLASSIFIER (203 MFLOP S , 9.03 MS , 11.1 M J).

TABLE II A LLOCATION OF A FIXED REPORT LENGTH Lz = nc T AT k = 3, γs = γr = 20 D B; THE TOKEN COUNT IS T = Lz /nc . M EANS OVER TEN RUNS , SAMPLE SD IN POINTS IN PARENTHESES . B OLD : BEST MEAN PER LENGTH . A DASH MARKS A FACTORIZATION OUTSIDE THE SWEPT GRID .

semantic width nc k MFLOPs

%

time [ms] (IQR) % energy [mJ]

1 2 3 4 5

50 77 90 97 100

2.29 (2.0–2.5) 3.91 (3.6–4.3) 5.20 (5.0–5.7) 6.86 (6.6–8.2) 8.63 (8.2–10.4)

101 156 183 196 203

mean test accuracy

ρ = 32 ρ = 128 ρ = 256

25 43 58 76 96

2.9 4.5 6.8 9.3 11.9

ρ = 512 ρ = 1024 C-AMC [5], ρ = 256

Lz

2

4

8

16

8 0.701 (1.4) 0.844 (1.4) 0.835 (2.6) – 16 0.745 (1.5) 0.864 (0.5) 0.876 (1.7) 0.863 (2.6) 32 – 0.869 (1.8) 0.889 (0.8) 0.896 (1.4)

local inference 0.92 85% of local

TABLE III F LAT R AYLEIGH BLOCK FADING AGAINST AWGN ON THE REPORTING LINK , AT k = 3, γs = 20 D B, AT THE MATCHED (nc , T ) SHOWN . M EANS OVER TEN RUNS , SAMPLE SD IN POINTS IN PARENTHESES .

γr = 20 dB 0.8

ρ

(nc , T )

AWGN

Rayleigh

γr = 0 dB AWGN

Rayleigh

32 (16, 4) 0.892 (0.9) 0.868 (3.0) 0.834 (2.5) 0.743 (1.4) 64 (8, 4) 0.889 (0.8) 0.877 (1.8) 0.783 (2.0) 0.676 (0.9) 256 (8, 1) 0.835 (2.6) 0.817 (1.3) 0.598 (3.0) 0.497 (1.7) 1024 (2, 1) 0.612 (1.5) 0.577 (1.7) 0.204 (1.0) 0.190 (0.9)

0.6 0.4 0.2 0

5

10

15

20

reporting SNR γr [dB] Fig. 3. Mean test accuracy versus reporting SNR γr at k = 3, γs = 20 dB, one curve per compression ratio ρ at its validation-selected (nc , T ). Open diamonds, dashed: C-AMC [5] at the matched ρ = 256. Error bars are ±1 sample SD.

Choosing k = 3 over k = 4 gives up the 0.7 and 1.6 points noted above, but saves 24% of node inference time and 27% of node energy. We therefore adopt k = 3 for the experiments that follow, as the better accuracy–cost compromise. b) Effect of semantic report length: Fig. 3 reports, for each compression ratio at k = 3, the mean accuracy of the factorization (nc , T ) selected on validation accuracy. At γr = 20 dB, accuracy degrades gradually with compression. At ρ = 256 it is 0.83, which retains 90% of the 0.92 local-inference accuracy. It remains above 0.80 through ρ = 512 and falls sharply to 0.61 at ρ = 1024. This degradation depends strongly on the reporting channel: the shorter the report, the more accuracy is lost as the channel worsens. Decreasing γr from 20 to 0 dB costs eight percentage points at ρ = 32 but 41 points at ρ = 512, from 0.82 to 0.41. At γr = 0 dB, accuracy has already fallen to 0.60 at ρ = 256. Shorter reports leave fewer transmitted dimensions to withstand channel distortion. Table II isolates the allocation of a fixed report length Lz = nc T between semantic width and temporal resolution. Allocation matters most where the budget is scarcest: the spread across factorizations is 13 to 14 percentage points at Lz ≤ 16 but only 2.7 points at Lz = 32. The loss is driven by insufficient width rather than by insufficient temporal resolution. The nc = 2 configurations do not exceed 0.75 at any evaluated token count, whereas at Lz ≤ 16 every configuration with nc ≥ 4 lies within 1.3 points of the best at the same budget. A minimum semantic width is therefore

necessary, above which the division between width and time is not critical. c) Flat fading on the reporting link: Under flat Rayleigh block fading the Lz real uses are paired into Lz /2 complex c c c symbols zk,i , received as zek,i = h zk,i + ηic , where a single gain h ∼ CN (0, 1) is drawn per frame and ηic ∼ CN (0, 2σr2 ). We assume that the edge server has perfect CSI. It applies the c c regularized linear equalizer zbk,i = h∗ zek,i /(|h|2 + 2σr2 ), and the real and imaginary parts of the result form b zk . The gain |h|2 scales the instantaneous per-use SNR while leaving its average unchanged. The report is assumed to lie within one coherence time and bandwidth. Table III retrains the same configurations through this link. At γr = 20 dB, the penalty relative to AWGN is only 1.2 to 3.5 percentage points, but it grows to 10.7 points at γr = 0 dB. Fading thus amplifies the sensitivity to the reporting SNR: the accuracy lost between γr = 20 and 0 dB is 1.9 times larger under fading at ρ = 64 and 2.1 times larger at ρ = 32. At ρ = 1024, the two channels lose accuracy at nearly the same rate, because performance there is limited by the bottleneck rather than by the link. d) Comparison with basic placements and the state of the art: Table IV sets the proposed split against the two basic placements of Sec. III and against C-AMC, all at the matched report length Lz = 32. All four schemes are trained on the same data, with the same augmentation, training budget and reporting-SNR sampling. The three split rows are the same backbone cut at different depths. C-AMC keeps its published architecture. Its total size is similar, 0.27 M parameters against our 0.28 M, but 93% of them sit at the edge server. The C-AMC embedding is placed under our perframe power normalization and reporting channel. We report the variant obtained without its published dropout stage, which we measured to reduce inference-mode accuracy substantially at our frame length. This choice favors the baseline. The intermediate split is more accurate than both basic

5

TABLE IV C OMPARISON AT Lz = 32 (ρ = 64), γs = γr = 20 D B. S PLIT ROWS USE (nc , T ) = (8, 4) AND THE RUNS OF F IG . 2; INDEPENDENT RUNS OF THE PROPOSED CONFIGURATION GIVE 0.889 (TABLE II). M EANS OVER TEN RUNS , SAMPLE SD IN POINTS IN PARENTHESES .

Node Scheme

par.

Edge

MFLOPs

par.

Acc.

k = 1, min. depth 0.05 M k = 5, max. depth 0.27 M C-AMC [5] 0.02 M

101 203 36

0.24 M 0.01 M 0.25 M

0.766 (0.4) 0.847 (2.0) 0.759 (2.5)

Proposed, k = 3

183

0.12 M 0.892 (1.2)

0.16 M

TABLE V M EAN ACCURACY VERSUS SENSING SNR AT k = 3, Lz = 16, WITH BOTH MODELS TRAINED OVER γs ∈ [−6, 20] D B. S AMPLE SD IN POINTS IN PARENTHESES .

Split γs [dB]

Local

20 10 5 0

0.820 (2.5) 0.581 (1.8) 0.284 (3.5) 0.119 (0.9)

γr = 20 dB

γr = 0 dB

0.721 (4.9) 0.530 (5.5) 0.264 (3.4) 0.101 (1.0)

0.525 (3.8) 0.368 (3.2) 0.209 (2.8) 0.099 (2.4)

placements. Reporting from the first stage has the lowest node cost but is 12.6 points below k = 3. Reporting from the last stage requires 68% more node parameters and 66% more node time than k = 3 and remains 4.5 points below it. The node cost of C-AMC is lower still, 0.02 M parameters against 0.16 M for the proposed split. The comparison is therefore a trade-off rather than a dominance: C-AMC trails by 13.3 points at this rate. Its accuracy is close to that of our k = 1 placement (0.759 against 0.766), which also pairs a small node model with a large edge model. The C-AMC gap is therefore comparable to the 12.6 points gained from k = 1 to k = 3, and much larger than the difference between k = 3 and k = 4 on the plateau of Fig. 2. In Fig. 3, C-AMC at ρ = 256 (dashed) lies below our curve at every reporting SNR. At γr = 20 dB, the gap stays between 13.2 and 16.0 points at every report length from Lz = 2 to 64. e) Sensing SNR and class-set size: To evaluate lower sensing SNR, we retrained the k = 3 split at Lz = 16 and the local classifier over γs ∈ [−6, 20] dB (Table V). This wider training range lowers the accuracy of both models at high γs . The split retains about 90% of the local accuracy down to γs = 5 dB. The loss caused by a poor reporting link shrinks as γs falls, because the observation itself becomes the limiting factor. At γs = 0 dB, both models are close to chance. We also repeated the comparison of Table IV with a reduced set of 8 classes and an expanded set of 24. With 24 classes, the ranking of the split depths is unchanged, and its margins stay within one point of the 16-class values. With 8 classes, all placements approach the local accuracy on a good reporting link, and the split depth matters only when the link is poor. V. C ONCLUSION This letter decoupled network split depth from semantic report length in remote AMR. It characterized the resulting

trade-off among sensing-node complexity, reporting cost and recognition accuracy over a noisy reporting channel. Intermediate splits provided the best trade-off, and an eight-use report retained 90% of local-inference accuracy. At a matched reporting rate, the intermediate split outperformed both basic placements and a state-of-the-art collaborative scheme. Running the whole backbone at the sensing node had the highest node cost and was still less accurate, because no edge-side capacity remained to process the noisy report. The intermediate split kept its advantage with 24 classes and retained about 90% of local accuracy at lower sensing SNR. Increasing observation bandwidth raises encoder computation, while adaptive pooling keeps the report length fixed by construction. Multi-antenna observation, faster channel variation, frequency-selective reporting links and class-set mismatch are left for future work. R EFERENCES [1] H. Zhang, F. Zhou, H. Du, Q. Wu, and C. Yuen, “Revolution of wireless signal recognition for 6G: Recent advances, challenges and future directions,” IEEE Commun. Surveys Tuts., vol. 28, pp. 3521–3563, 2026. [2] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based radio signal classification,” IEEE J. Sel. Topics Signal Process., vol. 12, no. 1, pp. 168–179, 2018. [3] D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 5–41, 2023. [4] J. Shao and J. Zhang, “BottleNet++: An end-to-end approach for feature compression in device-edge co-inference systems,” in Proc. IEEE ICC Workshops, 2020, pp. 1–6. [5] P. Dong, C. He, S. Gao, F. Zhou, and Q. Wu, “Edge-learning-based collaborative automatic modulation classification for hierarchical cognitive radio networks,” IEEE Internet Things J., vol. 11, no. 21, pp. 34 443–34 454, 2024. [6] S. Peng, S. Sun, and Y.-D. Yao, “A survey of modulation classification using deep learning: Signal representation and data preprocessing,” IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 12, pp. 7020–7038, 2022. [7] H. Wang, Z. Qi, Z. Li, and X. Zhao, “MR-Transformer: FPGA accelerated deep learning attention model for modulation recognition,” IEEE Trans. Wireless Commun., vol. 24, no. 2, pp. 1221–1233, 2025. [8] Y. Wang, G. Gui, H. Gacanin, B. Adebisi, H. Sari, and F. Adachi, “Federated learning for automatic modulation classification under class imbalance and varying noise condition,” IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 1, pp. 86–96, 2022. [9] Y. Matsubara, M. Levorato, and F. Restuccia, “Split computing and early exiting for deep learning applications: Survey and research challenges,” ACM Comput. Surv., vol. 55, no. 5, pp. 1–30, 2023. [10] J. Park, S. Oh, and S.-L. Kim, “SplitAMC: Split learning for robust automatic modulation classification,” in Proc. IEEE VTC2023-Spring, 2023, pp. 1–6. [11] C. Chaccour, W. Saad, M. Debbah, Z. Han, and H. V. Poor, “Less data, more knowledge: Building next-generation semantic communication networks,” IEEE Commun. Surveys Tuts., vol. 27, no. 1, pp. 37–76, 2025. [12] C. Wang, R. Olsson, S. Forsström, and Q. He, “Deep semantic inference over the air: An efficient task-oriented communication system,” in Proc. IEEE WCNC, 2026, pp. 1–6. [13] J. Shao, Y. Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck approach,” IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 197–211, 2022. [14] E. Bourtsoulatze, D. Burth Kurka, and D. Gündüz, “Deep joint sourcechannel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, 2019. [15] L. Boegner, M. Gulati, G. Vanhoy, P. Vallance, B. Comar, S. KokaljFilipovic, C. Lennon, and R. D. Miller, “Large scale radio frequency signal classification,” arXiv:2207.09918, 2022.

Related documents

• Semantic Split Inference for Remote Modulation Recognition#1217057
• Adaptive Unequal Error Protection for Semantic Split Learning over Wireless Channels#620634
• Adaptive Unequal Error Protection for Semantic Split Learning over Wireless Channels#618955
• KAIROS: Stateful, Context-Aware Power-Efficient Agentic Inference Serving#120489
• Multi-Source and Multi-Dimensional Information Fusion in Resource-Constrained Edge Networks#1297650