RepTC: Representation-Aware Optimization for Efficient Traffic Classification on Edge IoT Devices

Adel Chehade; Edoardo Ragusa; Paolo Gastaldo; Rodolfo Zunino

cs.NIOct 3, 2026

RepTC: Representation-Aware Optimization for Efficient Traffic Classification on Edge IoT Devices

Authors: Adel Chehade, Edoardo Ragusa, Paolo Gastaldo, Rodolfo Zunino

Organizations: Department of Naval, Electrical, Electronic and Telecommunications Engineering (DITEN), University of Genoa, 16126 Genoa, Italy

Abstract

Traffic classification (TC) is crucial to secure Internet of Things (IoT) networks, whose edge nodes often operate under privacy, bandwidth, and energy constraints. Yet, encrypted payloads and limited computing power make accurate, real-time TC a challenging task. Existing learning-based TC approaches often fix the input configuration a priori, even though it directly influences both predictive performance and computational cost. This paper presents RepTC, a representation-aware hardware-constrained strategy that addresses session-level TC through joint model and input optimization. The method co-optimizes network architecture, session length, and header preprocessing; this enables joint control of model complexity, input scale, and data representation within a unified resource-constrained design space. The proposed approach enforces microcontroller-class constraints on memory, model size, and computation, and yields compact models deployable on low-power edge devices. A gateway monitors traffic by aggregating sessions and either performs inference locally or offloads to a low-power edge node. Both scenarios are validated on heterogeneous embedded hardware, including a Raspberry Pi 3B+, STM32 Nucleo-F401RE, and XIAO ESP32-C3, spanning Cortex-A, Cortex-M, and RISC-V processor architectures; measured inference latency ranges from 0.63 to 18.59 ms, with MCU-side inference energy between 0.50 and 1.41 mJ per session. RepTC yields configurations that achieve high accuracy on a variety of established benchmarks: 96.21% on ISCX VPN-nonVPN, 99.62% on USTC-TFC2016, and 99.97% on Edge-IIoTset; at the same time, model size and computational requirements were reduced by up to three orders of magnitude compared with state-of-the-art methods. The results show that representation-aware optimization can improve efficiency while preserving competitive classification performance.

Figures & tables

		Architecture Design		Input Representation
Work	Gran.	Model	Design Strategy	Header Preprocessing	Header Fields Used	Input Shape	Input Size Selection	Hardware Aware
[ 8 ]	Packet	CNN/MLP	Search	IP anonym.	Fixed	600B	Fixed	✗
[ 13 ]	Packet	1D-CNN	Manual	Eth. removal + IP anonym. + UDP zero-pad	Fixed	1500B	Fixed	✗
[ 1 ]	Flow	Cascaded BLN	Manual	MAC/IP/port zero	Fixed	784B	Fixed	✗
[ 7 ]	Flow	1D-CNN	Search	None reported	Fixed	784B	Fixed	✗
[ 20 ]	Flow	CapsNet	Manual	IP zero	Fixed	784B	Fixed	✗

TABLE II : Comparison of representative raw-traffic-based TC approaches with respect to traffic granularity, architecture design, input representation, and hardware awareness.

Operation	S1	S2	S3	S4	S5	S6	S7	S8
Eth. Rem.	✓	✓	✓	✓
MAC Zero					✓	✓	✓	✓
IP Anon.	✓	✓			✓	✓
IP Zero			✓	✓			✓	✓
Port Zero	✓		✓		✓		✓

TABLE III : Preprocessing strategies defined in

\mathcal{S}_{h}

(a_{p},S_{h}^{p},L_{s}^{p})=\argmax_{c=1,\dots,N_{c}}E(a_{c},S_{h}^{c},L_{s}^{c},\mathcal{X_{V}})

Algorithm 1 RepTC: Joint Search over Architecture, Preprocessing, and Input Size

Fig. 1 : Class distribution overview for all evaluation datasets.

Fig. 2 : Cumulative session length distribution across datasets.

#	Layer	Input Dim	Filt/Units	Ker/Pool	Str	Pad
RepTC-ISCX model (ISCX VPN-NonVPN dataset)
Params: 54.67k , M-Tens: 2.55k , FLOPs: 305.87k , $\mathbf{S_{h}=4}$ , $\mathbf{L_{s}=218}$
1	Conv1D + ReLU	218 x 1	69	7	6	same
2	Conv1D + ReLU	37 x 69	61	4	7	valid
3	Conv1D + ReLU	5 x 61	48	1	1	same
4	Conv1D + ReLU	5 x 48	132	5	5	same

TABLE V : Best architectures identified by RepTC

Fig. 3 : Mutual information and Pearson correlation of search-space factors with validation accuracy.

Fig. 4 : ANOVA $p$ -value and $\eta^{2}$ effect-size heatmaps for key architectural and input factors.

Method	Input	Acc. (%)	F1 W (%)	F1 m (%)	Params. (M)	M-Tens. (K)	FLOPs (M)
ISCX VPNNon-VPN Dataset
RepTC-ISCX	sess.	96.21 $\pm$ 0.51	96.24 $\pm$ 0.51	95.33 $\pm$ 0.48	0.05	2.55	0.30
[ 6 ]	Sess.	96.60	96.63	-	0.09	20.12	10.08
[ 44 ]	Sess.	96.59	96.54	-	0.09	20.12	10.08
[ 46 ]	Sess.	-	91.80	-	0.22	313.60	267.22
[ 52 ]	Sess.	99.79	99.80	-	2.06	200.70	3152.27

TABLE VI : Comparison with representative prior works in terms of reported predictive performance and architecture-level complexity indicators.

Setting	Search Over			Results						Wall- clock (h)
	$a$	$S_{h}$	$L_{s}$	Acc. (%)	F1 m (%)	F1 w (%)	Par. (M)	M-Tens. (K)	FLOPs (M)
Restricted search ablations
Architecture search	✓	$S_{2}$	784 B	92.74 $\pm$ 0.76	92.33 $\pm$ 0.58	92.79 $\pm$ 0.74	0.03	3.02	0.50	17.76
Architecture+ $S_{h}$ search	✓	✓	784 B	93.12 $\pm$ 0.77	92.45 $\pm$ 0.67	93.15 $\pm$ 0.75	0.03	2.46	0.26	16.76
Architecture+ $L_{s}$ search	✓	$S_{2}$	✓	93.39 $\pm$ 0.80	92.69 $\pm$ 0.74	93.44 $\pm$ 0.78	0.04	1.12	0.46	17.26
Decoupled $R\!\rightarrow\!a$ search

TABLE VII : Search ablation results.

Fig. 5 : Convergence of the evaluated search configurations.

Level	Limit	$S_{h}$	$L_{s}$	Acc. (%)	F1 m (%)	F1 w (%)	Par. (M)	M-Tens. (K)	FLOPs (M)
Maximum tensor size / RAM
Constrained	2K	$S_{4}$	328 B	93.65 $\pm$ 0.21	92.55 $\pm$ 0.45	93.44 $\pm$ 0.25	0.04	1.27	0.26
Relaxed	8K	$S_{2}$	742 B	96.12 $\pm$ 0.32	95.54 $\pm$ 0.65	96.10 $\pm$ 0.27	0.04	4.27	0.46
Computational cost / FLOPs
Constrained	0.25M	$S_{4}$	234 B	92.65 $\pm$ 0.42	92.55 $\pm$ 0.61	92.44 $\pm$ 0.33	0.05	0.91	0.24
Relaxed	1.00M	$S_{1}$	236 B	96.65 $\pm$ 0.20	96.05 $\pm$ 0.55	96.54 $\pm$ 0.17	0.06	2.80	0.76

TABLE VIII : Sensitivity to individual resource constraints.

Fig. 6 : Accuracy impact of preprocessing strategy per dataset.

Fig. 7 : Recall of Edge-IIoTset DDoS classes across $S_{h}$ .

Fig. 8 : Resource–accuracy tradeoffs across input lengths.

Specification	Raspberry Pi 3B+	STM32 Nucleo- F401RE	XIAO ESP32-C3
CPU/Core	Cortex-A53	Cortex-M4	RISC-V
Clock	4 $\times$ 1.4 GHz	84 MHz	160 MHz
RAM	1 GB LPDDR2	96 KB SRAM	400 KB SRAM
Flash / Storage	microSD (ext.)	512 KB Flash	4 MB Flash
OS / SDK	Raspbian Lite	STM32Cube.AI	ESP-IDF / TFLM
Connectivity	Ethernet, UART	UART	USB/UART

TABLE XIII : Deployment platform specifications.

Fig. 9 : Quantization performance of RepTC models.

Fig. 10 : Deployment metrics across models and scenarios.

Edge Inference

Edge inference runs trained AI models on nearby devices such as phones, cameras, robots, and embedded computers under tight resource and latency limits.

See all 228 papers

Jul 22, 2026

Self-organizing Architecture of Receptron Units: a Hardware-Aware Framework for Edge Intelligence

The growing demand for intelligent processing at the edge of IoT networks is constrained by the severe computational and memory limitations of microcontroller units, which render impractical conventional deep learning approaches. We propose a neuromorphicinspired classifier based on the Receptron model, a single-unit architecture capable of implementing non-linearly separable decision boundaries, without resorting to multi-layer networks. The model is designed for direct deployment on mid-range MCUs, while supporting continuous on-device adaptation. Experimental evaluation on basic dataset benchmarks yields cross-validated accuracies compatible with standard machine learning method baselines. These results position the Receptron as a viable and interpretable alternative for resource-constrained neuromorphic edge systems operating in dynamic, non-stationary environments.

Stefano Radice, Ludovico Casaccia, Riccaro Emanuele Beccalli +2

Department of Physics “Aldo Pontremoli” and CIMaINa University of Milano Milano, Italy

Shallow Neural Networks Neuromorphic Computing Edge ML

Ask PDF

Aug 22, 2026

Pruned Traffic Trees: Native Semantic Compression with a Protocol-Structured Model Family for Encrypted Traffic Classification

Deep learning has achieved strong performance in encrypted traffic classification (ETC), yet its computational cost limits deployment on resource-constrained network devices such as routers and middleboxes. Existing compression methods mainly operate on weights, channels, hidden representations, or predictions, but do not explicitly determine which protocol fields and structural contexts should remain. We propose Pruned Traffic Trees (PTT), a three-level protocol-structured model family that treats native protocol structures as compression units. PTT-Full learns protocol-structured representations and field salience from complete Protocol Tree Graphs (PTGs), with flow-level self-supervised learning and protocol-presence-aware sparse execution. The learned salience and TopK+

k

closure construct Distilled PTGs (PTG-Ds) for PTT-Distilled, while PTT-Lite inherits this topology and reduces width through structure-aligned transfer and flow-level logits distillation. Under flow-disjoint and Strong Information Information (SII)-masked settings, PTT-Full achieves Macro-F1 scores of 0.9519 and 0.9416 on CSTNET-TLS1.3 and CipherSpectrum, while PTT-Lite retains 0.9325 and 0.9136 with 80.3% and 61.3% fewer parameters, 98.85% and 98.78% lower effective GFLOPs, and 8.75

\times

and 8.46

\times

CPU inference speedups. These results demonstrate that treating protocol structure itself as the compression object enables effective performance-efficiency trade-offs for lightweight ETC.

Yuantu Luo, Jun Tao, Xiangyu Xu +2

School of Cyber Science and Engineering, Southeast University, Nanjing, China · School of Computer Science and Engineering, Southeast University, Nanjing, China

Encrypted Traffic Classification Encrypted Network Traffic Analysis Model Compression Neural Network Compression

Ask PDF

Sep 14, 2026

Adaptive AI: Energy Efficient Multi-exit TinyML on Intelligent Vision Systems at the Edge

Traditional TinyML systems for edge devices achieve high accuracy by relying on fixed-depth models that require a constant number of multiply-accumulate (MAC) operations regardless of the input complexity. This approach wastes critical resources in battery-powered Internet-of-Things (IoT) devices and limits the real-time performance of edge cyber-physical systems. Multi-exit execution schemes mitigate these issues and are widely used on high-end devices such as GPUs, but are rarely exploited on edge IoT devices because they require substantial rethinking given their strict memory and computational constraints. We address these aspects by designing and deploying, on an ultra-low-power GWT GAP9 System-on-Chip (SoC), a novel multi-exit computational scheme, demonstrating it on a MobileNetV2 convolutional neural network (CNN) for the ImageNet-100 classification task. Our approach introduces multiple exits at different CNN depths, each with a confidence-based gating mechanism that dynamically and autonomously decides whether to continue or stop inference. Comparing our multi-exit strategy to the standard MobileNetV2 on a GAP9 SoC, we show a 41% reduction in the average computational cost (from 313 MMAC to 185 MMAC), a 29% lower inference time (from 49 to 35 ms), and an energy saving of 24% (from 2.1 to 1.6 mJ per frame). All these improvements come with a ~1% loss in accuracy compared to the full-depth MobileNetV2, which achieves 80.5%. Finally, comparing our adaptable multi-exit scheme with a third-party state-of-the-art adaptive CNN, also deployed on the GAP9, we achieve more than 2x its computational efficiency, increasing it from 8.1 to 17.2 MAC/cycle.

Luca Crupi, Lorenzo Lamberti, Alessandro Giusti +1

IDSIA, SUPSI, Lugano, Switzerland · IIS, ETH Zürich, Switzerland

Efficient Neural Network Inference Energy-Efficient ML Edge Inference Tiny ML Adaptive Computation Early-Exit Neural Networks

Ask PDF

Abstract

Edge Inference

Explore similar work

Self-organizing Architecture of Receptron Units: a Hardware-Aware Framework for Edge Intelligence

Pruned Traffic Trees: Native Semantic Compression with a Protocol-Structured Model Family for Encrypted Traffic Classification

Adaptive AI: Energy Efficient Multi-exit TinyML on Intelligent Vision Systems at the Edge