Towards Predicting Efficient and Anonymous Tor Circuits
summary
PredicTor uses a Random Forest trained on active probes and is PredicTor implemented in Tor source and shipped via the consensus; Shadow and live Tor tests show Random Forest predicts fast vs slow circuits at ~70–76% and yields 11–23% faster median downloads vs Vanilla and 7–13% vs the prior best, with a 500 ms median and >1.5 s 90th percentile speedup in Shadow, 680 km shorter circuits and 85% relay utilization; CLASI reveals similar sender‑AS leakage to Vanilla while CAR leaks more.
AI-written summary · about
chapters (18)
- 0:00Start
- 0:29Anonymity with Tor
- 1:24Latency in Tor
- 1:56Tor Congestion
- 2:18Path Selection Algorithms
- 2:39Bandwidth Weighted Path Selection
- 3:07Snader and Borisov
- 4:02Congestion Aware Routing
- 5:46Feature Extraction
- 6:34Evaluating Accuracy
- 8:41Shadow Experiment
- 9:20Live Tor Experiment
- 10:49Key Findings
- 12:05Predictor Performance Gains
- 13:15Client AS Inference (CLASI)
- 15:01Predictor Security Evaluation CLASI
- 16:16Conclusion
- 16:46Questions?
related talks
31mDeTor: Provably Avoiding Geographic Regions in Tor
54mAnonymity Loves Resilience: The Case of Tor
19mValidaTor: Domain Validation over Tor
20mDemystifying the Darknet in 30mns or Less
01h 16mAnonymity loves Diversity: The Case of Tor
44mAdvancing Human Rights With Tor
20mSocksTrace: A Proxy Leak Detector for Anonymity-Focused Network Applications
29mIdentifying and Characterizing Sybils in the Tor Network