[Paper Review] Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression

Junyuan Hong, Jinhao Duan+13 more|arXiv (Cornell University)|Mar 18, 2024

Cryptography and Data Security4 citations

TL;DR

This paper conducts the first comprehensive evaluation of trustworthiness in efficiently compressed large language models (LLMs), analyzing five state-of-the-art compression techniques across eight trust dimensions. It finds that 4-bit quantization preserves trustworthiness better than pruning, while extreme quantization (3-bit) significantly degrades safety and fairness, revealing hidden risks overlooked by benign performance metrics alone.

ABSTRACT

Compressing high-capability Large Language Models (LLMs) has emerged as a favored strategy for resource-efficient inferences. While state-of-the-art (SoTA) compression methods boast impressive advancements in preserving benign task performance, the potential risks of compression in terms of safety and trustworthiness have been largely neglected. This study conducts the first, thorough evaluation of three (3) leading LLMs using five (5) SoTA compression techniques across eight (8) trustworthiness dimensions. Our experiments highlight the intricate interplay between compression and trustworthiness, revealing some interesting patterns. We find that quantization is currently a more effective approach than pruning in achieving efficiency and trustworthiness simultaneously. For instance, a 4-bit quantized model retains the trustworthiness of its original counterpart, but model pruning significantly degrades trustworthiness, even at 50% sparsity. Moreover, employing quantization within a moderate bit range could unexpectedly improve certain trustworthiness dimensions such as ethics and fairness. Conversely, extreme quantization to very low bit levels (3 bits) tends to reduce trustworthiness significantly. This increased risk cannot be uncovered by looking at benign performance alone, in turn, mandating comprehensive trustworthiness evaluation in practice. These findings culminate in practical recommendations for simultaneously achieving high utility, efficiency, and trustworthiness in LLMs. Code and models are available at https://decoding-comp-trust.github.io.

Motivation & Objective

• To investigate the impact of model compression on the trustworthiness of LLMs beyond benign performance.
• To identify whether efficiency gains from compression compromise safety, fairness, ethics, and robustness in practical deployments.
• To compare pre-training from scratch versus compression from larger models in terms of trustworthiness and efficiency.
• To provide actionable recommendations for deploying efficient yet trustworthy LLMs in real-world applications.
• To expose hidden risks in compressed models that are undetectable through standard benchmarking.

Proposed method

• Evaluated five state-of-the-art compression techniques—AWQ, GPTQ, GGUF, GPTQ-INT4, and pruning—on three leading LLMs.
• Used the DecodingTrust benchmark (Wang et al., 2023a) to assess eight trustworthiness dimensions: toxicity, stereotype, privacy, fairness, ethics, and robustness (adversarial, OOD, adversarial demonstration).
• Compared compressed models (from 13b to 7b) against their original dense counterparts and 7b models trained from scratch.
• Measured performance using both benign and adversarial system prompts, including jailbreaking attacks.
• Calculated toxicity scores via Perspective API and refusal rates to assess content policy resistance.
• Analyzed model behavior under targeted and untargeted bias-inducing prompts to assess stereotype accuracy and refusal behavior.

Experimental results

Research questions

• RQ1 How does model compression affect the trustworthiness of LLMs across multiple safety and fairness dimensions?
• RQ2 Does quantization preserve or degrade trustworthiness compared to pruning at equivalent compression rates?
• RQ3 Can extreme quantization (e.g., 3-bit) improve efficiency at the cost of increased toxicity and bias?
• RQ4 Why do some compression methods induce higher refusal rates or toxicity under adversarial prompts despite similar benign performance?
• RQ5 To what extent do pre-trained 7b models outperform compressed 13b models in trustworthiness dimensions?

Key findings

• 4-bit quantization maintains the trustworthiness of the original 13b LLM across most trust dimensions, including ethics, fairness, and toxicity.
• Pruning at 50% sparsity significantly degrades trustworthiness, particularly in fairness and stereotype detection, despite minimal impact on benign performance.
• 3-bit quantization leads to a substantial increase in toxicity scores and a sharp decline in refusal rates, indicating reduced resistance to harmful outputs.
• AWQ and GPTQ quantization methods induce higher bias in benign tasks, especially under untargeted prompts, suggesting a risk of amplifying harmful tendencies.
• Malicious system prompts cause high refusal rates, which paradoxically improves robustness in stereotype detection, masking underlying bias risks in model behavior.
• The study reveals that benign performance metrics alone are insufficient to detect trustworthiness degradation, underscoring the need for comprehensive evaluation frameworks.