Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·23h ago
AI Research
arXiv CS.AI·May 28·5 min read
SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance-Diversity Data Selection
TL;DR
Fine-tuning machine learning models can lead to harmful outcomes if the training data is biased or unsafe. SPARD, a new defense mechanism, utilizes safety projection and relevance-diversity data selection to mitigate these risks.
✦ Why It Matters
Engineers can implement SPARD to enhance the safety of their fine-tuned machine learning models effectively.
Key Takeaways
- Adopt SPARD to defend against harmful fine-tuning attacks.
- Utilize Relevance-Diversity data selection for better safety coverage.
- Implement safety projections to maintain model integrity.
- Evaluate model performance on datasets like GSM8K and OpenBookQA.
How It Works
SPARD operates by alternating between optimizing the model's utility and applying safety projections. The safety projections are derived from a curated dataset that is selected using a Relevance-Diversity Determinantal Point Process, which ensures that the data is both relevant to the task and diverse enough to cover various safety aspects.