An Evolutionary Optimization Approach to Bit-Flip Attacks on LLMs

PDF view of the paper entitled Genbfa: The Evolutionary improvement approach to bitten extinction attacks on LLMS, by Sanjay Das and 5 other authors

PDF HTML (experimental) view

a summary:LLMS models have revolutionized the treatment of natural language (NLP), which excels in tasks such as generating text and summary. However, their increasing dependence on important important applications raises concerns about devices based on devices, especially extinction attacks (BFAS). BFAS, enabling rift injection methods such as Rowhammer, the parameters of the target model in memory, exposing both integrity and performance. Determining the critical parameters of BFAS in the space of the vast parameter of LLMS constitute great challenges. While previous research indicates that the transformer -based structure is by its nature more powerful in BFAs compared to traditional deep nerve networks, we challenge this assumption. For the first time, we make it clear that at least three points can cause the deterioration of catastrophic performance in LLM with billions of parameters. Current BFA technologies are not sufficient to exploit this weakness due to the difficulty of identifying critical parameters efficiently within the enormous teacher space. To address this, we suggest attention, which is a new LLMS framework that allows an effective crossing of the teacher’s space to determine critical parameters. In addition, we offer Genbfa, a development strategy that improves an evolutionary designed to improve research more, and isolate the most important bits of effective and effective attack. Experimental results reveal the depth weak from LLMS to attention. For example, just three bits (4.129 x 10^-9 % of the total parameters) in the Llama3-8B-Instruct 8-bit model (W8) resulting in a complete collapse of performance: accuracy in MMLU tasks from 67.3 % to 0 %, and Wikitext Perplexity Skyrocelles of 12.6 Lux 4.7. These results emphasize the effectiveness of attention to detecting and exploiting security gaps within LLM structures.