Abstract
In response to the growing popularity of Tor and other censorship circumvention systems, censors in non-democratic countries have increased their technical capabilities and can now recognize and block network traffic generated by these systems on a nationwide scale. New censorship-resistant communication systems such as SkypeMorph, StegoTorus, and CensorSpoofer aim to evade censors' observations by imitating common protocols like Skype and HTTP. We demonstrate that these systems completely fail to achieve unobservability. Even a very weak, local censor can easily distinguish their traffic from the imitated protocols. We show dozens of passive and active methods that recognize even a single imitated session, without any need to correlate multiple network flows or perform sophisticated traffic analysis. We enumerate the requirements that a censorship-resistant system must satisfy to successfully mimic another protocol and conclude that "unobservability by imitation" is a fundamentally flawed approach. We then present our recommendations for the design of unobservable communication systems.
Introduction
As national firewalls, such as the Great Firewall of China, evolve from simple IP blocking to real-time Deep Packet Inspection (DPI), a new class of "parrot" circumvention systems has emerged. Tools like SkypeMorph, StegoTorus, and CensorSpoofer attempt to achieve unobservability by wrapping forbidden traffic in a layer that mimics the statistical properties—such as packet sizes and timings—of benign protocols like Skype video calls or web browsing. This research demonstrates that this "imitation game" is an asymmetric battle that circumvention tools are fundamentally losing.
Methodology
The study establishes a taxonomy of adversaries, ranging from a Local Adversary (LO) (e.g., a coffee shop Wi-Fi administrator) to a State-level Omniscient Adversary (OM) with massive processing power. The researchers subjected leading parrot systems to two primary types of analysis:
- Passive Attacks: Examining traffic for "tells," such as missing side protocols (e.g., Skype’s TCP control channel) or incorrect metadata in supposedly benign files (e.g., malformed PDF headers in StegoTorus).
- Active/Proactive Attacks: Actively perturbing the network—dropping packets or closing connections—to observe if the parrot reacts like the genuine protocol. For instance, a real Skype client increases control traffic when it detects packet loss; a parrot that only mimics the data stream fails this dynamic response.
Experiments were conducted on a live testbed using official implementations of the circumvention tools compared against genuine instances of Skype, SIP (Ekiga), and various web servers (Apache, IIS).
The Flaws of Protocol Mimicry
The research identifies several critical insights that make imitation a failing strategy:
- The Asymmetry of Perfection: To succeed, a parrot must be perfect in every dimension—packet headers, timing, side protocols, and even implementation-specific bugs. A censor only needs to find a single discrepancy to block the traffic.
- Side Protocols as the Smoking Gun: Most parrots focus on the main data channel but ignore mandatory "side-cars" like TCP control messages, STUN for NAT traversal, or HTTP update checks.
- Dynamic Dependencies: Real protocols are reactive. If a censor drops packets and the "Skype call" does not attempt to renegotiate its codec or increase control traffic, it is identified as a fake.
- Implementation Quirks: Mimicking a protocol is insufficient; one must mimic a specific version on a specific OS. Differences in how Windows versus Linux handles TCP sequence numbers can unmask a parrot.
Experimental Results
The findings represent a "total break" for the analyzed systems. Both SkypeMorph and StegoTorus failed basic passive tests, lacking mandatory fields in UDP headers and failing to generate expected TCP control traffic.
In active testing, the systems showed a complete lack of the dynamic reactions expected from the protocols they were imitating. For example, when researchers dropped UDP packets or closed TCP channels, the parrot systems exhibited "no reaction," whereas a real Skype client would have initiated probes or shifted ports.
Even in a "Parrot+" experiment, where authors attempted to fix these systems by adding fake control channels, the tools still failed because the fake channels were not logically correlated with the data channel in the way a real application’s internal logic would dictate.
Conclusion and Future Directions
The researchers conclude that imitation is a dead end. The complexity of perfectly mimicking a modern protocol is too high to be sustainable. Instead of synthesizing fake traffic, the authors suggest that circumvention systems should run the actual protocol.
Systems like FreeWave or SWEET represent a more promising path: tunneling. By using a genuine application (like a real Skype client) to carry encrypted data within its audio/video payloads, the censor is forced into a difficult choice: either break the encryption of a legitimate, popular application or block the application entirely—a decision with significant political and economic costs.
Summary updated