AI POWERED DETECTION OF ADVERSARIAL AND SUPPLY CHAIN ATTACKS ON GENERATIVE MODELS

Abstract

Generative models, including Generative Adversarial Networks (GANs) and diffusion-based architectures, have become mainstream in computer vision, synthetic data generation and digital media content generation. But the intricacy of their design and the nature of their data driven training pipelines make them susceptible to different types of advanced threats, in particular – adversarial attacks and supply chain attacks. In this paper we propose a state-of-the-art AI-based detection framework that is able to detect and take actions against Threats such as these in real time. In this work, we present a novel hybrid architecture which utilizes transformer-based anomaly detection in latent space in conjunction with cryptographic and structural analysis for the supply chain verification. In particular, the detection pipeline contains two complementary components: i) behavioral fingerprinting of generative outputs based on transformer encodings, where attention-weighted embeddings are examined to reveal potential latent inconsistencies suggesting adversarial tampering; and ii) integrity checking of model dependencies and weights via cryptographic hash chaining and dependency graph analysis. We validate our method with exhaustive experiments of high-fidelity generative models such as StyleGAN2 and Stable Diffusion with benchmark datasets of CIFAR-10 and FFHQ. Not only does this framework attain more than 94% in detection accuracy, but it also passes the precision of 91% with false positive rates almost completely lower than those of traditional anomaly detectors such as Isolation Forests and LSTM-Autoencoders. Our solution provides a scalable and secure foundation for protecting generative AI systems in critical infrastructures, offering the interpretability afforded by Explainable AI (XAI) without sacrificing either robust or high-resolution anomaly detection.