LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation

The introduction of LoopLLM marks a significant advancement in the field of large language models (LLMs), particularly in mitigating energy-latency attacks that exploit computational vulnerabilities. By focusing on repetitive generation, LoopLLM effectively induces low-entropy decoding loops, allowing LLMs to reach over 90% of their maximum output length, a stark improvement from the mere 20% achieved by previous methods. This innovation not only enhances the performance of LLMs but also improves their transferability across different models by approximately 40%. Such developments are vital as they address the growing concerns regarding the efficiency and security of LLMs in various applications, ensuring that these models can operate effectively under potential attack scenarios.