Attacking Compressed Vision Transformers

• DOI: 10.48550/arxiv.2209.13785
• Publication Date: 2022-01-01
• Authors: Swapnil Parekh; Devansh Shah; Pratyush Shukla
• Keywords: Edge device; Robustness

Vision Transformers are increasingly embedded in industrial systems due to their superior performance, but memory and power requirements make deploying them edge devices a challenging task. Hence, model compression techniques now widely used deploy models on as they decrease the resource inference very fast efficient. But reliability robustness from security perspective is another major issue safety-critical applications. Adversarial attacks like optical illusions for ML algorithms can severely impact accuracy of models. In this work we investigate transferability adversarial samples across SOTA Transformer 3 compressed versions infer effects different have attacks.