IDFuzz: Intelligent Directed Grey-box Fuzzing
summary
IDFuzz uses a neural network to learn which input offsets matter and then mutates those fields, claiming a 2.48x median speedup on the Google Fuzzer Test Suite and 91.86% fewer ineffective mutations; it finds six new vulnerabilities and yields four CVE IDs so far. Key techniques are branch encoding to represent sibling branches, adaptive data set generation to train compact models, and gradient filtering plus DB‑scan clustering to turn gradients into practical field mutations.
AI-written summary · about
chapters (10)
- 0:00IDFuzz and directive fuzzing
- 1:13Limitations of directive fuzzing
- 2:16Mutation strategy challenges
- 3:00Neural network model
- 4:04JPEG parsing vulnerability
- 5:49Gradient-guided mutation
- 6:55IDFuzz innovations
- 8:21Evaluation results
- 9:49Discovered vulnerabilities
- 10:11Conclusion
related talks
01h 00mClusterFuzz: Fuzzing at Google Scale
53mFuzzing 101
39mFuzzing Session: Finding Bugs and Vulnerabilities Automatically
31m“Fuzz or lose..."
13mWingFuzz: Implementing Continuous Fuzzing for DBMSs
01h 15mMake your programs more reliable with Fuzzing
34mODDFuzz: Hunting Java Deserialization Gadget Chains via Structure-Aware Directed Greybox Fuzzing
53mThree Colours of Fuzzing: Reflections and Open Challenges
