Combining Formal and Informal Information in Bayesian Program Analysis via Soft(…)
summary
Converts a Datalog analysis into a Bayesian network, attaches rule probabilities, and injects neural-network confidences as soft evidence (implemented via a noisy sensor t′ trick) rather than hard labels so ranking changes but soundness stays. Uses semantic embeddings from Graph CodeBoard and a small MLP to predict alias or intent-match probabilities. Evaluated on 8 Java and 8 Android benchmarks: pointer-analysis ranking improves by 55.4%/44.9%/58% (inversion/mean/median) and taint by 67.2%/44.7%/37.6%; artifacts and code are available.
AI-written summary · about
chapters (14)
- 0:00Combining formal and informal info
- 0:20Motivation for informal info
- 1:25Challenges in informal hints
- 2:10Pointer analysis example
- 3:56Bayesian analysis framework
- 5:20Limitations of hard evidence
- 6:22Soft evidence for uncertain hints
- 8:10Updating alarm probabilities
- 8:31Neural networks for alias prediction
- 9:14Full Bayesian-neural approach
- 9:37Evaluation on Java and Android
- 12:29Dynamic test integration
- 13:10Conclusion and future
- 13:52Q&A
related talks
18mLearning Abstraction Selection for Bayesian Program Analysis
01h 03mAn Overview of Probabilistic Programming
32mStructured Deep Learning with Probabilistic Neural Prog
41mScalable Collective Inference from Richly Structured Data using Probabilistic Soft Logic (PSL)
18mActive Learning for Neurosymbolic Program Synthesis
24mFuzzing Guided by Bayesian Program Analysis
40mProbabilistic Programming for Augmented Intelligence
39mThe Duct Tape of Heroes: Bayes Rule