A new framework for classifying encrypted network traffic

The FLARE framework uses field-aware learning to classify data even when encryption hides most information.

A new framework called FLARE has been developed for encrypted traffic classification. This framework is designed to identify different types of network traffic. It works even when encryption makes much of the data difficult to see.

Network security monitoring and traffic management require the ability to classify traffic. This process helps with service identification and detecting malicious activity. It also supports quality-of-service assurance.

In older versions of the TLS 1.2 protocol, observers could see plaintext handshake information. This included the server certificate and the Server Name Indication (SNI). These details helped identify the service being used.

Modern protocols like TLS 1.3 and QUIC change this. TLS 1.3 encrypts the server certificate and messages following the ServerHello. The Encrypted Client Hello (ECH) mechanism further hides the SNI and other sensitive parameters.

QUIC also protects much of the transport-control information. It applies header protection to packet numbers and certain header bits. These changes mean that observable bytes are mostly ciphertext. This makes it hard to link byte patterns to application semantics.

The FLARE framework uses a method called Masked Field Prediction (MFP). This process involves a BERT-based packet encoder. The encoder reconstructs masked tokens using the surrounding field-structured context.

Field-aware pre-training means the system learns from the specific organization of protocol fields. This helps the encoder understand field values and how they depend on each other. It focuses on information that remains observable during encryption.

During the fine-tuning stage, the parameters of the pre-trained PacketEncoder remain frozen. A packet-level gated module then combines protocol-semantic representations with directional packet-length representations. This happens at the packet level.

A flow-level encoder then models the dependencies between different packets. This allows the system to perform classification for specific tasks. The framework adapts to different scenarios by combining these different types of information.

The researchers tested FLARE on eight different datasets. These included the CipherSpectrum and VisQUIC datasets. These datasets represent scenarios with very strong encryption.

FLARE showed superior performance compared to other baseline models. On the QUIC website fingerprinting task, it exceeded the strongest pre-trained baseline by 10.92 percentage points. This shows it works well even when observability is constrained.

The framework uses a packet-level gated module to merge different data types. This module combines semantic information with the direction and length of packets. This helps the system handle varying levels of available information.

Key facts

  • FLARE outperformed the strongest pre-trained baseline by 10.92 percentage points on QUIC website fingerprinting.
  • The framework uses Masked Field Prediction to reconstruct tokens from field-structured context.
  • The system was evaluated using eight different datasets, including CipherSpectrum and VisQUIC.
  • FLARE uses a frozen PacketEncoder to produce protocol-semantic representations during fine-tuning.

All sections

Edition