T-BFA: Targeted Bit-Flip Adversarial Weight Attack.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep Neural Networks (DNNs) are vulnerable to adversarial attacks.
• Existing attacks primarily focus on input examples, not weight parameters.
• Bit-Flip based adversarial weight Attack (BFA) is a powerful method targeting DNN weights.

Purpose of the Study:

• To propose the first targeted Bit-Flip based adversarial weight attack (T-BFA) for DNNs.
• To demonstrate the ability to intentionally mislead selected inputs to a specific target output class.
• To evaluate the effectiveness of T-BFA on various DNN architectures for image classification.

Main Methods:

• Developed a class-dependent weight bit searching algorithm to identify critical weight bits.
• Implemented T-BFA to manipulate a small number of weight bits in DNN models.
• Tested T-BFA on DNN architectures like ResNet-18 using the ImageNet dataset.

Main Results:

• Achieved 100% attack success rate in misclassifying 'Hen' images into 'Goose' class using ResNet-18.
• Manipulated only 27 out of 88 million weight bits for the attack.
• Maintained a validation accuracy of 59.35% on the affected model.

Conclusions:

• T-BFA is a highly effective method for targeted adversarial weight attacks on DNNs.
• The attack demonstrates significant potential for compromising DNN integrity with minimal modifications.
• This research highlights a new, potent vulnerability in DNN security.