Abstract
Internet censorship has evolved from static keyword blocking to
dynamic, adaptive deep packet inspection (DPI) and machine learning-based
filtering, capable of adapting to protocol obfuscation in near real time. Traditional circumvention tools, such as VPNs and Tor, often fail in these
situations. We propose an adaptive, feedback-driven packet manipulation
framework that dynamically selects and switches circumvention strategies in
real time. Our system integrates multi-signal censorship detection, a decision
engine optimized using a multi-armed bandit, and a pluggable segmentation
subsystem that supports multiple sharding strategies. New features include a pivot scheduling scheme with failover and cooldown periods to prevent oscillation, and a comprehensive benchmarking framework for recovery time, adaptation frequency, and throughput stability. In experiments conducted at
CensorLab, our Segmenters achieved an 83% circumvention success rate and maintained throughput under high load, outperforming both fixed and random
segmentation baselines. These results demonstrate that feedback-driven multi- strategy adaptation offers a promising path to resilient censorship
circumvention in the face of an evolving threat landscape.
dynamic, adaptive deep packet inspection (DPI) and machine learning-based
filtering, capable of adapting to protocol obfuscation in near real time. Traditional circumvention tools, such as VPNs and Tor, often fail in these
situations. We propose an adaptive, feedback-driven packet manipulation
framework that dynamically selects and switches circumvention strategies in
real time. Our system integrates multi-signal censorship detection, a decision
engine optimized using a multi-armed bandit, and a pluggable segmentation
subsystem that supports multiple sharding strategies. New features include a pivot scheduling scheme with failover and cooldown periods to prevent oscillation, and a comprehensive benchmarking framework for recovery time, adaptation frequency, and throughput stability. In experiments conducted at
CensorLab, our Segmenters achieved an 83% circumvention success rate and maintained throughput under high load, outperforming both fixed and random
segmentation baselines. These results demonstrate that feedback-driven multi- strategy adaptation offers a promising path to resilient censorship
circumvention in the face of an evolving threat landscape.
| Original language | English |
|---|---|
| Title of host publication | Proceedings of the International Conference on Smart Technologies and Reliable Systems (SmartRS25) |
| Publisher | Springer Nature |
| Publication status | Accepted for publication - 28 Oct 2025 |
| Event | International Conference on Smart Technologies and Reliable Systems: SmarTRS 2025 - Najran, Saudi Arabia Duration: 28 Oct 2025 → 29 Oct 2025 https://smartrs.net/ |
Publication series
| Name | Lecture Notes on Data Engineering and Communications Technologies |
|---|---|
| Publisher | Springer Nature |
| ISSN (Print) | 2367-4512 |
| ISSN (Electronic) | 2367-4520 |
Conference
| Conference | International Conference on Smart Technologies and Reliable Systems |
|---|---|
| Country/Territory | Saudi Arabia |
| City | Najran |
| Period | 28/10/25 → 29/10/25 |
| Internet address |
Keywords
- Censorship circumvention
- Packet segmentation
- Feedback-driven evasion
- Traffic Shaping
- Network resillience