SemPurify: Semantics-Aware Data Purification for Black-box Backdoor Defense

Key points

• Backdoor attacks have emerged as a critical security threat in machine learning, where adversaries embed triggers within poisoned training samples to manipulate a model's predictions on targeted inputs while leaving its general behavior unaffected.1
• Current purification-based defenses use image transformations (e.g., blurring) to disrupt trigger patterns but often degrade image quality.1
• Using this semantic information, the T2I model subsequently refines the transformed samples, helping preserve visual details and semantic information during restoration.1
• Extensive experiments spanning multiple backdoor scenarios, i.e., classification models, CLIP-style encoders, and generative models, demonstrate that SemPurify reliably removes backdoors while effectively preserving semantic information for downstream tasks.1

Sources (1)

• [1]SemPurify: Semantics-Aware Data Purification for Black-box Backdoor Defense
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 03:09 PM

  To address this, we propose the first semantics-aware data purification framework for black-box backdoor defense, SemPurify, which explicitly incorporates multi-level semantics (i.e., low-level semantic embeddings and high-level image tags) within a text-to-image (T2I) framework to guide image restoration during purification.

  Backdoor attacks have emerged as a critical security threat in machine learning, where adversaries embed triggers within poisoned training samples to manipulate a model's predictions on targeted inputs while leaving its general behavior unaffected.

Extractive summary: sentences quoted from the sources.