A framework for mitigating malicious RLHF feedback in LLM training using consensus based reward

Large Language models (LLMs) have demonstrated impressive capabilities in natural language processing and understanding. LLMs are being rapidly adopted in major industry sectors including mobile computing, healthcare, finance, government, and education driven by technology giants such as NVIDIA, OpenAI, Microsoft, Apple, Meta, Google, Broadcom, AMD, and IBM. However, due to the emerging nature of this technology, many security/privacy challenges remain unresolved that we must tackle before rolling out LLMs to critical applications (e.g. Healthcare, Legal). In this article, we focus on the Reinforcement Learning via Human Feedback (RLHF) process that is widely used for training LLMs giving them the human-like feel most applications value. The RLHF process involves employing human experts to generate feedback based on an LLM's query-response pairs and using this feedback to then retrain (fine-tune) the model. However, RLHF can also expose the LLM to malicious feedback generated by one or more individuals in the process leading to degraded performance of the LLM and harmful responses. Most state-of-the-art (SOTA) solutions to this problem involve utilizing a KL-Divergence-based brute-force update-rejection approach that can render the whole RLHF process completely useless (model quality is not improved) in the presence of malicious entities in the process. We propose the COnsensus-Based RewArd framework (COBRA), a consensus-based technique that can effectively negate the malicious noise generated by a certain segment of the RLHF human-expert pool, leading to improved LLM training performance in a mixed-trust scenario. We have evaluated COBRA  for two separate LLM use cases, Sentiment Analysis and Conversational Task. We have experimented with a wide range of LLM models (e.g. GPT-2 XL - 1.5B parameters). COBRA  outperformed the standard unprotected reward generation scheme by [Formula: see text] for the generative conversational task and by [Formula: see text] for the sentiment analysis task. We have also quantitatively compared COBRA  with Coste et al. and observed state-of-the-art performance, particularly when a lower number of reward models are used ([Formula: see text] increased reward accuracy at [Formula: see text]).