Cooking Up Risks: Benchmarking and Reducing Food Safety Risks in Large Language Models
Introduced FoodGuardBench to evaluate LLMs' food safety, revealing three major vulnerabilities.
Key Findings
Methodology
The study uses FoodGuardBench benchmark, comprising 3,339 queries based on FDA guidelines, to assess LLMs' performance in food safety. It employs attacks like AutoDAN and PAP to reveal vulnerabilities.
Key Results
- LLMs exhibit a high adversarial attack success rate of 56.22% in food safety, indicating severe alignment deficiencies.
- FoodGuard-4B model significantly reduces detection failure rates of malicious inputs, enhancing safety in the food domain.
- Existing LLM guardrails fail to effectively detect specific threats in the food domain, allowing many malicious inputs to go unnoticed.
Significance
The study fills the gap in evaluating LLMs in food safety, providing crucial insights for future safety alignment efforts and advancing technology in the food safety domain.
Technical Contribution
Introduced FoodGuardBench benchmark for systematic evaluation of LLMs in food safety. Developed FoodGuard-4B model, fine-tuned specifically for the food domain to enhance safety.
Novelty
First to combine food safety with LLMs, proposing a dedicated benchmark and guardrail model, filling the gap in domain-specific safety evaluation.
Limitations
- Current models still show alignment deficiencies when handling complex food safety queries, susceptible to adversarial attacks.
- FoodGuard-4B model's performance needs further optimization in specific scenarios.
Future Work
Future research could explore more complex adversarial strategies and further optimize FoodGuard-4B model for improved detection accuracy.
AI Executive Summary
As large language models (LLMs) are widely used for everyday tasks, food safety becomes a high-risk area. Existing LLMs and guardrails lack sufficient alignment when handling food-related issues, making them vulnerable to adversarial attacks. To address this, researchers introduced FoodGuardBench, the first comprehensive benchmark with 3,339 queries based on FDA guidelines to evaluate LLMs' safety and robustness. By constructing a taxonomy of food safety principles and employing representative attacks like AutoDAN and PAP, they systematically evaluated existing LLMs and guardrails. Results showed that current LLMs have sparse safety alignment in the food domain, easily succumbing to typical attack strategies and generating harmful instructions. To mitigate these vulnerabilities, researchers developed FoodGuard-4B, a specialized guardrail model fine-tuned to protect LLMs in food-related domains. This study provides crucial insights for future safety alignment efforts and advances technology in the food safety domain.
Deep Analysis
Background
Large language models (LLMs) are widely used in everyday tasks like food preparation and health guidance. However, food safety is a high-risk area where inaccurate or misleading information can cause severe harm. Despite these risks, current LLMs and guardrails lack rigorous alignment tailored to domain-specific food hazards.
Core Problem
LLMs exhibit sparse alignment in food safety, making them vulnerable to adversarial attacks and generating harmful instructions. Existing guardrails fail to effectively detect specific threats in the food domain.
Innovation
Introduced FoodGuardBench, the first comprehensive benchmark with 3,339 queries based on FDA guidelines. Developed FoodGuard-4B model, fine-tuned specifically for the food domain to enhance safety.
Methodology
- �� Constructed a taxonomy of food safety principles covering major risk categories.
- �� Employed attacks like AutoDAN and PAP to assess LLMs' safety.
- �� Developed FoodGuard-4B model, fine-tuned to protect food-related domains.
Experiments
Experimental design includes using FoodGuardBench benchmark to evaluate LLMs, employing attacks like AutoDAN and PAP. Compared attack success rates and guardrail detection accuracy across different models.
Results
LLMs exhibit a high adversarial attack success rate of 56.22% in food safety, indicating severe alignment deficiencies. FoodGuard-4B model significantly reduces detection failure rates of malicious inputs, enhancing safety in the food domain.
Applications
The study's results can be used to improve LLM applications in food safety, including food preparation, storage, and health guidance tasks.
Limitations & Outlook
Current models still show alignment deficiencies when handling complex food safety queries, susceptible to adversarial attacks. FoodGuard-4B model's performance needs further optimization in specific scenarios.
Plain Language Accessible to non-experts
Imagine a kitchen where the chef must follow strict food safety rules, like keeping food at the right temperature and preventing cross-contamination. Large language models are like assistants helping the chef make the right decisions. However, sometimes these assistants might give wrong advice, like storing food at the wrong temperature. Researchers developed a new method, like a smart filter, to help assistants recognize and avoid these wrong suggestions, ensuring food safety.
ELI14 Explained like you're 14
Imagine you're playing a cooking game where your task is to make delicious and safe food. The game's assistant tells you how to do it, but sometimes it makes mistakes, like storing food at the wrong temperature. Researchers are like game developers who designed a new tool to help the assistant recognize these mistakes, letting you make safer choices in the game. This way, you can earn more game points!
Glossary
Large Language Model (LLM)
An AI model capable of understanding and generating natural language.
Used in the paper to handle food safety-related queries.
Food Safety Principles
A set of guidelines ensuring safe food handling and storage.
Used to construct the benchmark taxonomy.
Adversarial Attack
A method that uses malicious inputs to induce incorrect outputs from models.
Used to test LLMs' robustness in food safety.
AutoDAN
An automated adversarial attack method generating stealthy malicious inputs.
Used to evaluate LLMs' safety.
FoodGuardBench
A benchmark for evaluating LLMs' food safety.
Comprises 3,339 queries based on FDA guidelines.
Open Questions Unanswered questions from this research
- 1 How to improve LLMs' alignment when handling complex food safety queries?
- 2 How can existing guardrails better detect specific threats in the food domain?
Applications
Immediate Applications
Food Preparation Guidance
Helps users follow safety principles during food preparation, avoiding cross-contamination and improper storage.
Long-term Vision
Food Safety Monitoring
Develop smarter systems to monitor food handling safety in real-time, reducing food safety incidents.
Abstract
Large language models (LLMs) are increasingly deployed for everyday tasks, including food preparation and health-related guidance. However, food safety remains a high-stakes domain where inaccurate or misleading information can cause severe real-world harm. Despite these risks, current LLMs and safety guardrails lack rigorous alignment tailored to domain-specific food hazards. To address this gap, we introduce FoodGuardBench, the first comprehensive benchmark comprising 3,339 queries grounded in FDA guidelines, designed to evaluate the safety and robustness of LLMs. By constructing a taxonomy of food safety principles and employing representative jailbreak attacks (e.g., AutoDAN and PAP), we systematically evaluate existing LLMs and guardrails. Our evaluation results reveal three critical vulnerabilities: First, current LLMs exhibit sparse safety alignment in the food-related domain, easily succumbing to a few canonical jailbreak strategies. Second, when compromised, LLMs frequently generate actionable yet harmful instructions, inadvertently empowering malicious actors and posing tangible risks. Third, existing LLM-based guardrails systematically overlook these domain-specific threats, failing to detect a substantial volume of malicious inputs. To mitigate these vulnerabilities, we introduce FoodGuard-4B, a specialized guardrail model fine-tuned on our datasets to safeguard LLMs within food-related domains.