Nan Deng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Nan Deng has a total of 10 co-authors. Here are the top co-authors by number of shared publications:
- Feng Hong
Shanghai Acoustics Laboratory, Chinese Academy of Sciences, Shanghai 201805, China
1 shared publication - Hang Lei
University of Electronic Science and Technology of China, Chengdu, Sichuan, China
1 shared publication - Haozhe Zhou
School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China
1 shared publication - Jian Cheng
Research Institute of Mine Artificial Intelligence, Chinese Institute of Coal Science, Beijing, China
1 shared publication - Kehui Liu
Institute of Urban Systems Engineering, Beijing Academy of Science and Technology, Beijing, China
1 shared publication - Maolin Yang
University of Electronic Science and Technology of China, Chengdu, Sichuan, China
1 shared publication
Latest Publications
- Open Access
- 29 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 120-127
SilentDNS: Acoustic Side-Channel-Assisted Inference for Device Fingerprinting and State Detection under Encrypted Network Traffic
The pervasive adoption of Transport Layer Security (TLS) 1.3, QUIC (UDP-based multiplexed secure transport), and Encrypted Client Hello (ECH) renders over 95% of network traffic opaque to deep packet inspection, fundamentally challenging device fingerprinting and state detection in enterprise and home networks. Existing encrypted traffic analysis methods relying solely on network metadata degrade by more than 30% under multi-device concurrency, traffic shaping defenses, and noisy labels, exposing a critical gap: the absence of an orthogonal information source that persists independently of encryption protocol evolution. This paper presents SilentDNS, a cross-modal inference framework that fuses ambient acoustic side-channel signals with encrypted network traffic metadata for device fingerprinting and operational state detection without decryption. Unlike prior acoustic side-channel work that treats sound as an attack vector, SilentDNS is the first to leverage device acoustic emissions as an auxiliary defensive channel, aligned with network observations through a weak-supervision InfoNCE contrastive mechanism requiring only coarse window-level labels. Evaluated across 12 device types in three environments, SilentDNS achieves 96.3% macro-F1 in device fingerprinting (+8.7 pp over the best network-only baseline) and 93.1% in state detection, with gains of up to 22.4 pp on I/O-intensive events. Under state-of-the-art traffic padding defenses, SilentDNS recovers 71% of the accuracy lost by network-only classifiers, demonstrating that cross-modal physical–digital fusion provides an observation channel fundamentally harder for adversaries to suppress than network-layer features alone. SilentDNS, a cross-modal framework that fuses ambient acoustic side-channel signals with encrypted network traffic metadata for device fingerprinting and state detection without decryption. SilentDNS exploits the observation that devices emit characteristic acoustic signatures (fan speed, hard disk drive (HDD) seeks, relay clicks) temporally correlated with network behavior—signals that persist regardless of encryption evolution. A contrastive alignment mechanism operates under weak, window-level supervision, and an integrated explanation engine generates human-readable evidence chains for SOC workflows. Across 12 device types in three environments, SilentDNS achieves 96.3% macro-F1 for fingerprinting (+8.7 pp over the best network-only baseline) and 93.1% for state detection (+22.4 pp on I/O-intensive events). Under traffic padding defenses, SilentDNS recovers 71% of accuracy lost by network-only classifiers.
- 29 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 120-127
Open AccessSilentDNS: Acoustic Side-Channel-Assisted Inference for Device Fingerprinting and State Detection under Encrypted Network Traffic
The pervasive adoption of Transport Layer Security (TLS) 1.3, QUIC (UDP-based multiplexed secure transport), and Encrypted Client Hello (ECH) renders over 95% of network traffic opaque to deep packet inspection, fundamentally challenging device fingerprinting and state detection in enterprise and home networks. Existing encrypted traffic analysis methods relying solely on network metadata degrade by more than 30% under multi-device concurrency, traffic shaping defenses, and noisy labels, exposing a critical gap: the absence of an orthogonal information source that persists independently of encryption protocol evolution. This paper presents SilentDNS, a cross-modal inference framework that fuses ambient acoustic side-channel signals with encrypted network traffic metadata for device fingerprinting and operational state detection without decryption. Unlike prior acoustic side-channel work that treats sound as an attack vector, SilentDNS is the first to leverage device acoustic emissions as an auxiliary defensive channel, aligned with network observations through a weak-supervision InfoNCE contrastive mechanism requiring only coarse window-level labels. Evaluated across 12 device types in three environments, SilentDNS achieves 96.3% macro-F1 in device fingerprinting (+8.7 pp over the best network-only baseline) and 93.1% in state detection, with gains of up to 22.4 pp on I/O-intensive events. Under state-of-the-art traffic padding defenses, SilentDNS recovers 71% of the accuracy lost by network-only classifiers, demonstrating that cross-modal physical–digital fusion provides an observation channel fundamentally harder for adversaries to suppress than network-layer features alone. SilentDNS, a cross-modal framework that fuses ambient acoustic side-channel signals with encrypted network traffic metadata for device fingerprinting and state detection without decryption. SilentDNS exploits the observation that devices emit characteristic acoustic signatures (fan speed, hard disk drive (HDD) seeks, relay clicks) temporally correlated with network behavior—signals that persist regardless of encryption evolution. A contrastive alignment mechanism operates under weak, window-level supervision, and an integrated explanation engine generates human-readable evidence chains for SOC workflows. Across 12 device types in three environments, SilentDNS achieves 96.3% macro-F1 for fingerprinting (+8.7 pp over the best network-only baseline) and 93.1% for state detection (+22.4 pp on I/O-intensive events). Under traffic padding defenses, SilentDNS recovers 71% of accuracy lost by network-only classifiers.
- Open Access
- 21 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 17 (19391404) , 7773-7780
Subsurface Defect Detection in GPR Data Integrating Temporal and Spatial Features
Ground Penetrating Radar (GPR) has emerged as a pivotal tool for subsurface explorations, particularly in detecting subsurface defects that might endanger structural integrity. While GPR B-scan data visually depict underground conditions, it represents the time delay and amplitude of the returned Electromagnetic (EM) waves, making them complex to interpret due to both their image- like appearance and their inherent waveform changes. To address this complexity, this paper introduces the novel Temporal-Spatial Synthesis Network (TSSNet), designed to harness both temporal and spatial features for enhanced subsurface defect detection in GPR B-scan data. The Echo State Network (ESN), underpinned by Reservoir Computing, is utilized to fit the GPR data and capture its “temporal features”, emphasizing the temporal variations present in the EM waves. Concurrently, the Convolutional Neural Network (CNN) focuses on discerning “spatial features” from the B-scan images, spotlighting spatial patterns that possibly indicate subsurface defects. After extracting these temporal and spatial features, they are synthesized to form a comprehensive representation of the GPR data. The enhanced synthesized feature facilitates precise classification, resulting in heightened differentiation between normal and defect-contained subsurface areas. Experiments on real-world GPR datasets are conducted, with the results underscoring the efficacy of the proposed approach.
- 21 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 17 (19391404) , 7773-7780
Open AccessSubsurface Defect Detection in GPR Data Integrating Temporal and Spatial Features
Ground Penetrating Radar (GPR) has emerged as a pivotal tool for subsurface explorations, particularly in detecting subsurface defects that might endanger structural integrity. While GPR B-scan data visually depict underground conditions, it represents the time delay and amplitude of the returned Electromagnetic (EM) waves, making them complex to interpret due to both their image- like appearance and their inherent waveform changes. To address this complexity, this paper introduces the novel Temporal-Spatial Synthesis Network (TSSNet), designed to harness both temporal and spatial features for enhanced subsurface defect detection in GPR B-scan data. The Echo State Network (ESN), underpinned by Reservoir Computing, is utilized to fit the GPR data and capture its “temporal features”, emphasizing the temporal variations present in the EM waves. Concurrently, the Convolutional Neural Network (CNN) focuses on discerning “spatial features” from the B-scan images, spotlighting spatial patterns that possibly indicate subsurface defects. After extracting these temporal and spatial features, they are synthesized to form a comprehensive representation of the GPR data. The enhanced synthesized feature facilitates precise classification, resulting in heightened differentiation between normal and defect-contained subsurface areas. Experiments on real-world GPR datasets are conducted, with the results underscoring the efficacy of the proposed approach.