Invention Description
Artificial intelligence, machine learning, and computer security are increasingly important as large language models (LLMs) and deep neural networks (DNNs) are used in a wide range of applications. However, these models are vulnerable to bit-flip attacks, which alter stored model parameters and can compromise their accuracy, reliability, and security. Existing attack methods may produce extreme parameter values, making them unstable and easier to detect, highlighting the need for more effective and stealthy approaches to evaluating AI model vulnerabilities.
Researchers at Arizona State University have developed a Sneaky Bit-Flip Attack (SBFA) framework which evaluates LLM vulnerability to bit-level memory faults using ImpactScore, which combines gradient sensitivity with constraints based on benign layer-wise weight distributions. It ranks candidate changes using gradient-based ImpactScore and accelerates the search with SKIP Search, an algorithm which prioritizes promising candidates while reducing the work required to find them. Experiments across Qwen, LLaMA, and Gemma models report substantial accuracy degradation from individual bit flips in BF16 and INT8 settings. Designed for authorized robustness research, SBFA can help teams assess model integrity across floating-point and quantized formats.
This novel framework identifies how a single carefully selected bit flip can severely degrade a large language model while keeping the altered weight within its layer’s normal value range.
Potential Applications
- AI security research and LLM robustness evaluation
- Red-team testing of deployed models and their memory-protection systems
- Development of fault detection, model integrity monitoring, and secure inference safeguards
- Assessment of hardware and software defenses for AI systems in cloud, enterprise, and edge environments
Benefits and Advantages
- Identifies high-impact bit-level vulnerabilities while constraining changes to layer-wise weight ranges
- Combines parameter sensitivity and perturbation magnitude to prioritize candidate changes
- Uses SKIP Search to reduce the computational burden of searching large models
- Supports evaluation of floating-point and quantized model formats
- Can reveal weaknesses that conventional testing or broader parameter perturbations may miss
- Provides a research framework for assessing model integrity and studying resilience to hardware-level faults
- Applies to both BF16 floating-point and INT8 quantized model weights
For more information about this opportunity, please see