5 April 2024 in Expert Systems with Applications
26 November 2024 in Computers & Security
6 March 2025 in Expert Systems with Applications
11 December 2024 in IEEE Transactions on Consumer Electronics
Edge AI-driven diffusion models (DMs) are increasingly integrated into consumer devices for high-quality data generation and content creation. This paper introduces InvisibleDiffusion, a novel backdoor attack framework for diffusion models in consumer electronics, designed to remain undetected by utilizing a non-standard Gaussian distribution as a concealed trigger. Unlike previous backdoor methods, InvisibleDiffusion does not rely on obvious visual triggers, enhancing its stealthiness. Extensive experiments demonstrate that InvisibleDiffusion achieves high attack efficacy against DDPM and DDIM models on CIFAR-10 and CelebA datasets, while maintaining the functional integrity of the models. Our code is available for reproducibility at https://anonymous.4open.science/r/b2hoaWNhbnRzZWV0aGF0bm9vb29vb29vb29v.
12 August 2025 in ACM Transactions on Multimedia Computing, Communications, and Applications
Many edge computing applications based on computer vision have harnessed the power of deep learning. As an emerging deep learning model for vision, Vision Transformer models have recently achieved record-breaking performance in various vision tasks. But many recent studies on the robustness of the Vision Transformer have shown that the Vision Transformer is still vulnerable to adversarial attacks and is easily affected by adversarial attacks, causing the model to misclassify the input. In this work, we ask an intriguing question: “Can Adversarial Perturbations against Vision Transformers be detected with model explanations?” Driven by this question, we observe that benign samples and adversarial examples have different attribution maps after applying the Grad-CAM interpretability method on the Vision Transformer model. We demonstrate that an adversarial example is a Feature Shift of the input data, which leads to an Attention Deviation of the visual model. We propose a framework for capturing the Attention Deviation of vision models to defend against adversarial attacks. Furthermore, experiments show that our model achieves expectative results.