MACS-BNet: A Stealthy Multiconstraint Adversarial Backdoor Network Against Compressed Learning

Deep learning-based compressed sensing techniques have exhibited exceptional prowess in signal reconstruction and data-sharing applications, particularly within the realm of IoT sensor data processing. However, existing methods overlook a critical security vulnerability: the susceptibility of compressed sensing techniques to backdoor attacks during the reconstruction phase, which could pose severe security risks to downstream applications. This study pioneers an investigation into the feasibility of backdoor injection during the reconstruction phase, presenting the stealthy multi-constraint adversarial backdoor network against compressed learning (MACS-BNet) and substantiating its efficacy in subverting downstream classification tasks. MACS-BNet synergistically incorporates detailed sensing enhancement, fortified by local information relative positional encoding (LiRPE), to elevate image reconstruction fidelity. Concurrently, it employs a multi-constrained adversarial optimization that integrates sparsity, amplitude regulation, and spatial smoothness constraints, achieving an optimal trade-off between perturbation imperceptibility and attack efficacy. Consequently, victim models are subtly manipulated to yield outputs consistent with the attacker’s objectives. Extensive empirical evaluations reveal that MACS-BNet consistently surpasses seven cutting-edge attack methodologies across attack success rate, clean sample classification accuracy, and stealthiness under both all-to-one and all-to-all attack paradigms. Specifically, MACS-BNet attains an unparalleled clean classification accuracy of 99.52% and an attack success rate of 99.43% in the all-to-one mode, while simultaneously ensuring high-quality image reconstruction. Furthermore, MACS-BNet exhibits formidable resistance against detection by seven state-of-the-art defense mechanisms, underscoring its superior stealth and robustness.