Anti-AdvTamp: exposing adversarial evasive tampering attacks in Network-on-Chips with a multi-scale attribute fusion detection

Abstract

Network-on-chip (NOC) architectures are increasingly becoming a widely adopted system-on-chip (SoC) architecture due to their scalable and modular communication structure, which matches the requirements of modern multiprocessor system-on-chip (MPSoC) and deep learning (DL) accelerator architectures, yet they are vulnerable to routing tampering attacks (e.g., traffic diversion) that induce Denial-of-Service (DoS). Therefore, machine learning (ML) have been increasingly applied to tampering detection in NoC; however, existing ML-based detectors struggle to cope with more covert adversarial tampering attacks. To address this challenge, this paper proposes SEAA, a symmetry-exploiting framework to construct two adversarial tampering attack models (AdvTamp1 and AdvTamp2), and evaluates their impact on ML-based detection. Furthermore, this research proposes a novel ML detection framework based on multi-scale attribute fusion, which integrates coarse-grained and fine-grained NoC performance metrics to improve detection accuracy. Experiments under PARSEC and synthetic Traffic Pattern benchmarks on various NoC topologies show that traditional ML models achieve detection accuracies of only 57.4% and 49.4% under AdvTamp1 and AdvTamp2 attacks, respectively. In contrast, our proposed multi-scale attribute fusion method significantly improves detection performance, achieving accuracies of 96.19% and 98% under AdvTamp1 and AdvTamp2 , respectively. These results demonstrate the effectiveness of the proposed method in improving detection under adversarial tampering and highlight its potential for enhancing the security of NoC.

More information

Identifiers

ORCID for Boojoong Kang: orcid.org/0000-0001-5984-9867

ORCID for Basel Halak: orcid.org/0000-0003-3470-7226

Catalogue record

Export record