Novel Cross-Channel Soft-XOR Adversarial Engine-based Advanced Autonomous Driving Security for Traffic Sign Recognition System

Abstract

AbstractDeep neural networks deployed in edge-computing perception layers are the backbone of modernautonomous driving systems. While the state-of-the-art YOLOv12 architecture family deliversunprecedented real-time object detection performance, its vulnerability to sophisticated adversarialmanipulation remains a critical security concern. Traditional gradient-based adversarial attacks,such as FGSM, BIM, and PGD, exhibit a severe trade-off between disruption capability and visualimperceptibility; they introduce highly visible, high-frequency linear noise patterns that are easilyintercepted by standard defensive filters, or they suffer from acute performance degradation whenfacing dense parameter spaces. To address this limitation, this paper introduces the Cross-ChannelSoft-XOR (CC-SXOR) adversarial engine, a novel math-driven framework designed to executehighly stealthy yet robust perception disruption. Operating through a non-linear color-spacecoupling loop (R → G → B → R), CC-SXOR leverages an algebraic continuous logical relaxationfunction that dynamically scales down perturbation amplitudes in uniform zones whilestrategically embedding adversarial vectors into high-contrast structural textures. We conduct acomprehensive, cross-scale, and cross-dataset evaluation targeting the full spectrum of the trainedYOLOv12 family—spanning from the lightweight Nano (n) configuration to the server-side ExtraLarge (x) variant. Benchmarked across three diverse, large-scale autonomous drivingenvironments—including the German Traffic Sign Recognition Benchmark (GTSRB), BDD100K,and the Udacity Self-Driving Car dataset—experimental results demonstrate that CC-SXORestablishes a superior fidelity-efficacy equilibrium. Crucially, while conventional baseline attacksundergo a severe operational collapse down to 0% Total Attack Success Rate (ASR) when facingmassive model capacities and complex scenes, the proposed CC-SXOR framework demonstratesscale-invariant transferability, consistently maintaining a stable disruption topology (ASR ≥ 50%).Concurrently, it preserves an unprecedented perceptual similarity profile, maintaining a PeakSignal-to-Noise Ratio (PSNR) of 34–35 dB and a Structural Similarity Index (SSIM) of 0.82–0.84across all scales, which represents a substantial gain of up to +7 dB in PSNR and +0.25 in SSIMover unconstrained state-of-the-art iterative frameworks like PGD. Finally, we provide anextensive structural ablation study and discuss the fundamental security implications of this nonlinear cross-channel vulnerability for emerging Edge AI vision architectures.Keywords: Cyber-Physical Security, Adversarial Machine Learning, Cross-Channel Perturbation,YOLOv12, Autonomous Driving Perception.