Abstract
Identifying application types within network traffic is essential for modern cybersecurity, yet traditional methods struggle with encrypted streams and dynamic port allocation. This paper introduces the Chained Graph Neural Network (CGNN), a novel approach that reimagines packet streams as "chained graphs" rather than simple sequences or images. By capturing the causal and structural relationships between packets using specialized Graph Neural Networks, CGNN achieves a 23% to 29% accuracy boost in application classification over traditional deep learning methods while maintaining high inference efficiency.
Introduction
The "old ways" of traffic classification—port-based and statistical methods—are increasingly ineffective against modern network techniques like Network Address Translation (NAT) and ubiquitous encryption. While researchers have turned to Deep Learning (DL), most existing models treat traffic as 2D images (CNNs) or text sequences (RNNs). The authors argue this creates a structure mismatch. Network traffic is a compositional process where the timing and order of packets carry vital information that standard DL models often fail to extract, leading to poor generalization.
Methods
The researchers propose a two-step pipeline to bridge this gap: transforming raw data into a graph structure and processing it through a specialized architecture.
Chained Graph Construction
Instead of feeding raw bytes directly into a neural network, a packet session is transformed into a graph. Each packet is treated as a vertex, represented by a 1500-byte feature vector. An undirected edge is then drawn between packets that appear adjacent in the stream, creating a "linear chain" graph that preserves the temporal flow of the conversation.
The CGNN Architecture
The model utilizes a four-layer structure designed for efficiency and structural awareness:
- Two SGC Layers: Simple Graph Convolution (SGC) is used for linear feature aggregation. The authors found that non-linear graph neurons performed worse as they disrupted the simple linear semantics of the chained graph.
- AVGPooling Layer: This aggregates vertex-level features into a single graph-level representation.
- Fully Connected Layer: A final layer maps these features to specific classification labels via a Softmax function.
By using SGC, the model remains computationally efficient while allowing each packet to "know" something about its predecessor and successor through feature smoothing.
Experimental Evaluation
The model was evaluated across three distinct traffic types:
- Application Traffic: 41 categories (e.g., web browsing, streaming) from a real-world testbed.
- Malicious Traffic: 5 malware types (e.g., Artemis, Emotet) to test security utility.
- Encrypted Traffic: The public ISCX VPN-nonVPN dataset, containing 11 categories of encrypted traffic.
Key Results
- Superior Accuracy: CGNN outperformed baselines like 1D-CNN and Deep Packet by 23% to 29% in application classification. In malware detection, accuracy improved by 2% to 37%.
- Encryption Robustness: The model reached approximately 95% accuracy on encrypted traffic. While 1D-CNN slightly outperformed it in specific VPN scenarios (98%), CGNN's focus on structural "fingerprints" makes it highly resilient when payloads are hidden.
- High Efficiency: CGNN's inference time is roughly 1 second, significantly faster than recurrent models like FS-Net (13 seconds) or Deep Packet (110 seconds).
Discussion and Insights
- The "Image" Fallacy: Treating pcap files as 2D images ignores the underlying logic of network protocols. Graphs provide a more "native" representation for non-Euclidean network data.
- Linearity as a Strength: Unlike many GNN applications where deep non-linear layers are preferred, linear aggregation (SGC) is superior for chained graphs. It captures the chain structure without the "over-smoothing" issues common in complex GNNs.
Conclusion and Future Work
While CGNN shows significant promise, limitations remain. The current model uses zero-padding for shorter packets, which can dilute features in datasets with small average packet lengths. Future work will focus on testing CGNN in "live" environments with tighter latency constraints and exploring more complex graph topologies that account for multi-flow interactions beyond simple chains.
Summary updated