Backdoor in Seconds: Unlocking Vulnerabilities in Large Pre-trained Models via Model Editing.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Computer Vision

Background:

• Large pre-trained models (e.g., ViT) are powerful but vulnerable to backdoor attacks.
• Existing attacks require access to training data and significant computational resources, posing challenges for large models.

Purpose of the Study:

• To investigate the unique challenges of backdoor attacks on large pre-trained models.
• To develop an effective and feasible backdoor attack method suitable for these models.

Main Methods:

• Introduced EDT (Efficient, Data-free, Training-free), a novel backdoor attack method.
• EDT utilizes model editing techniques by injecting a lightweight codebook to modify model behavior without data poisoning or retraining.

Main Results:

• Demonstrated the effectiveness of EDT across various pre-trained models (ViT, CLIP, BLIP, stable diffusion).
• Validated the attack's success on diverse downstream tasks like image classification, captioning, and generation.

Conclusions:

• EDT presents a viable solution for backdoor attacks on large pre-trained models, overcoming previous limitations.
• The method offers a practical approach to assessing the vulnerability of complex AI systems.