EdgeGuard: Unified Adversarial, Backdoor, and Side-Channel Defense for Edge AI Deployments

Read full text Summarise with VIRAT
Authors: Yuxuan Chen, Xinyi Li, Zihan Zhang, Junjie Wang, Chao Lu
Cited by: 0
References: 73

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Deploying deep learning models on edge devicessmartphones, IoT gateways, autonomous vehicles, and VR headsets-exposes them to a uniquely converged threat landscape where digital AI attacks and physical hardware vulnerabilities intersect. Adversarial examples manipulate model predictions, backdoor triggers embed hidden malicious behaviors, model extraction steals proprietary parameters, and physical side channels through charging interfaces, wireless links, and sensor peripherals leak private data about the model and its users. Existing defenses address these threats in isolation, leaving dangerous gaps that cross-domain attackers exploit. We present EDGEGUARD, a unified defense framework for edge AI that simultaneously protects against adversarial inputs, backdoor triggers, model theft, and physical-layer information leakage. EDGEGUARD introduces four tightly integrated components: (1) an Adaptive Input Sentinel (AIS) that detects adversarial and backdoor-triggered inputs through joint activation-distribution and gradient-saliency analysis without modifying the protected model; (2) a Side-Channel Isolation Engine (SCIE) that neutralizes electromagnetic, power, and acoustic information leakage during on-device inference by injecting hardware-aware obfuscation at the system-on-chip (SoC) level; (3) a Model Integrity Monitor (MIM) that detects and prevents extraction attacks through query-pattern analysis and output perturbation; and (4) a Federated Threat Sharing Protocol (FTSP) that enables privacy-preserving sharing of detected threat signatures across a fleet of edge devices. We evaluate EDGEGUARD on a testbed of 186 heterogeneous edge devices spanning 14 device models, testing against 8 distinct attack categories across 5 datasets and 4 model architectures over a 20-week deployment. EDGEGUARD achieves 96.8% overall threat detection rate at 1.9% false positive rate, reduces side-channel information leakage by 95.2%, and degrades only 4.1% under fully adaptive adversaries-with a total runtime overhead of 3.8 ms per inference and 5.3% battery consumption increase.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Security and Verification in Computing Physical Unclonable Functions (PUFs) and Hardware Security

Publication details

Publication date
23 March 2026
Language
EN
Document type
Preprint