Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Infusion can craft subtle edits to a small fraction (0.2%) of training data to induce targeted model behavior changes, competitive with inserting explicit behavior examples.result 90%
• Infusion uses scalable influence-function approximations to compute small perturbations to training documents that induce targeted changes in model behavior through parameter shifts.method 95%
• Infusion transfers across architectures (ResNet ↔ CNN), suggesting a single poisoned corpus can affect multiple independently trained models.result 90%
• Infusion is most effective at amplifying behaviors the model has already learned, rather than instilling entirely new ones.technical 85%
• The reliance on accurate influence function estimations, scalability to large datasets, and handling of discrete token spaces in language models pose challenges for Infusion.limitation 90%
• Small, subtle edits to training data can systematically shape model behavior, underscoring the importance of training data interpretability for adversaries and defenders.result 90%
• Infusion was evaluated on data poisoning tasks across vision and language domains, using CIFAR-10 and GPT-Neo models.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card