Biases in Edge Language Models: Detection, Analysis, and Mitigation

2025-02-17

AI Summary

This work identifies and mitigates the exacerbation of social bias in large language models (LLMs) when deployed on low-power edge devices—e.g., Raspberry Pi—where resource constraints degrade fairness. Experiments reveal that Llama-2 exhibits 43.23% and 21.89% higher fairness bias on Raspberry Pi 4 compared to desktop and cloud environments, respectively. To address this without retraining, we propose a layer-wise feedback correction mechanism: during inference, it dynamically constrains hidden-layer outputs via weighted regularization to suppress biased activations in real time. We further introduce a multi-environment fairness benchmark enabling temporal bias tracking and cross-platform evaluation. Evaluated across six models—including Raspberry Pi-optimized Llama-2, GPT-4o-mini, and Gemini-1.5-flash—the method achieves a 79.28% reduction in measured bias. This study is the first to systematically demonstrate the adverse impact of edge deployment on LLM fairness and establishes a plug-and-play, low-overhead governance paradigm for responsible lightweight AI deployment.

Technology Category

• Machine Learning: Ethics, Bias, and Fairness
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

The integration of large language models (LLMs) on low-power edge devices such as Raspberry Pi, known as edge language models (ELMs), has introduced opportunities for more personalized, secure, and low-latency language intelligence that is accessible to all. However, the resource constraints inherent in edge devices and the lack of robust ethical safeguards in language models raise significant concerns about fairness, accountability, and transparency in model output generation. This paper conducts a comparative analysis of text-based bias across language model deployments on edge, cloud, and desktop environments, aiming to evaluate how deployment settings influence model fairness. Specifically, we examined an optimized Llama-2 model running on a Raspberry Pi 4; GPT 4o-mini, Gemini-1.5-flash, and Grok-beta models running on cloud servers; and Gemma2 and Mistral models running on a MacOS desktop machine. Our results demonstrate that Llama-2 running on Raspberry Pi 4 is 43.23% and 21.89% more prone to showing bias over time compared to models running on the desktop and cloud-based environments. We also propose the implementation of a feedback loop, a mechanism that iteratively adjusts model behavior based on previous outputs, where predefined constraint weights are applied layer-by-layer during inference, allowing the model to correct bias patterns, resulting in 79.28% reduction in model bias.

Problem

• Detect bias in edge language models
• Analyze bias across deployment environments
• Mitigate bias using feedback loop mechanism

Innovation

• Optimized Llama-2 on Raspberry Pi
• Feedback loop for bias correction
• Layer-by-layer constraint weights application