Benchmarking Resilience: Asymmetric Latent Purification Inspired by Generative Diffusion Bottlenecks

Abstract

Deep learning classifiers exhibit susceptibility towards iterative adversarial perturbations, often under high-fidelity attacks experiencing total categorical collapse. To address this, we introduce the Asymmetric Latent Purifier (ALP), a novel structural defence mechanism inspired by the stochastic information bottlenecks of the 2026 Unified Latents (UL) generative framework, Unlike Traditional deterministic autoencoders, ALP incorporates an adaptive, non-differentiable Gaussian noise layer within a 64-channel latent manifold to disrupt adversarial gradient flows. Empirically validated on CIFAR-10 dataset using an Apple M4 8-core GPU architecture. While the unprotected baseline experiences a total categorical collapse ( 0.00% accuracy) under a 7-step iterative PGD attack, our 20-sample adaptive ensemble approach achieves a robust accuracy of 32.06% (SD=1.94%)( averaged over 5 trials ) while ensuring a high-fidelity reconstruction of 25.68 dB. Operating a total system latency of 13.86ms, offers a promising path towards real-time flexibility for complex RGB varieties. Furthermore, with a single-sample inference latency of 1.25 ms, ALP represents a 100x to 1000x speedup over iterative diffusion-based purifiers, enabling real-time adversarial immunity in safety-critical systems.