Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• SafePickle achieves 90.01% F1-score on their own dataset of 727 Pickle-based files from Hugging Face, significantly outperforming existing scanners (Modelscan, Fickling, ClamAV, VirusTotal) which achieve 7.23%-62.75%.result 95%
• SafePickle achieves 81.22% F1-score on out-of-distribution PickleBall data, outperforming the PickleBall method's 76.09% while remaining library-agnostic.result 95%
• SafePickle correctly classifies all 9/9 evasive Hide-and-Seek malicious models specially crafted to evade scanners, making it the only method to do so.result 95%
• SafePickle uses machine learning to statically extract structural and semantic features from Pickle bytecode without requiring per-library policy generation or code instrumentation.method 90%
• SafePickle is lightweight and generic, eliminating the need for complex system setups and verified benign models required by previous defenses like PickleBall.technical 85%
• The method could have limitations in handling new types of novel attacks not covered in existing datasets.limitation 80%
• False positives could potentially disrupt legitimate workflows if the system isn't tuned properly.limitation 80%
• SafePickle demonstrates that data-driven detection can effectively and generically mitigate Pickle-based model file attacks.result 85%

Competitive landscape

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

Startup potential card