Large-scale Mobile App Identification Using Deep Learning

Shahbaz Rezaei, Bryce Kroencke, Xin Liu

Abstract

Identifying network traffic is a cornerstone of modern network management, from security monitoring to billing enforcement. However, the rise of encryption and the proliferation of mobile apps have made traditional methods like Deep Packet Inspection (DPI) increasingly obsolete. This paper presents a deep learning framework for large-scale mobile app identification that operates effectively on encrypted traffic. By utilizing a 1D Convolutional Neural Network (CNN) on the initial bytes of a flow and a novel CNN+LSTM "flow association" technique for contextualizing ambiguous traffic, the proposed system achieves up to 98% accuracy across 80 popular apps.

Introduction

The shift toward ubiquitous encryption (SSL/TLS) and dynamic port allocation has created a "blind spot" for network operators. Traditional traffic classification relies on plain-text signatures or fixed ports, both of which are increasingly rare in mobile ecosystems. Furthermore, many apps generate "ambiguous flows"—background traffic for advertisements or analytics that look identical across different applications. If a network monitor sees a generic ad flow, it is traditionally impossible to tell if it originated from a game, a news app, or a navigation tool.

This research addresses these challenges by proposing a model that requires only the first few packets of a flow for classification, enabling early prediction for real-time applications like routing and Quality of Service (QoS) provisioning.

Methodology

The authors developed two primary architectures to handle different aspects of traffic classification:

1D CNN for Single-Flow Identification

To identify an app from a single stream of packets, the model processes the first 256 bytes of the header and payload from the first six packets. These are concatenated into a 1536-length vector and fed into a 1D CNN. The CNN's shift-invariance allows it to detect characteristic app signatures or TLS handshake patterns regardless of their exact position within the packet payload.

Standard CNN architecture used as a building block for the proposed identification models.

CNN+LSTM for Flow Association

To solve the problem of ambiguous flows, the authors introduced a "flow association" model. This architecture considers the target flow alongside adjacent flows occurring within a short temporal window (e.g., 2 seconds). Each flow is processed by the CNN to extract features, which are then sequenced through a Long Short-Term Memory (LSTM) network. This allows the model to learn the unique temporal "fingerprint" of connections an app makes upon execution, effectively using context to identify the source of generic background traffic.

Occlusion Analysis

To understand what the model "sees" in encrypted traffic, the researchers performed occlusion analysis. By systematically masking fields in the SSL/TLS handshake, they discovered that the CNN relies heavily on unencrypted fields such as the Server Name Indication (SNI), cipher suite lists, and custom extensions to differentiate between apps.

Experimental Results

The models were evaluated on a proprietary dataset of 80 popular mobile apps collected from a mobile operator network, including both iOS and Android traffic.

  • Clean Traffic: On traffic where ambiguous flows were removed, the CNN achieved 94.22% accuracy using only payload data. Including IP/port headers pushed this to 99.12%.
  • Ambiguous Traffic: When including background flows, the payload-only CNN accuracy dropped to 84.01%, which still represents a significant improvement over traditional methods at this scale.
  • Impact of Flow Association: For services with high ambiguity (like Google apps), the CNN+LSTM model improved accuracy from 91.98% to 96.23% by leveraging context from adjacent flows.
  • Robustness: The CNN+LSTM remained robust (~95% accuracy) even when unrelated "noise" flows from other apps were interleaved into the sequence.

Discussion and Conclusion

The study demonstrates that deep learning can effectively identify mobile apps even when traffic is encrypted, primarily by exploiting metadata leaked during the handshake process. The "flow association" technique is a significant advancement, providing a way to attribute generic background traffic to a specific source app by analyzing the sequence of network events.

However, the transition to TLS 1.3 poses a challenge, as it encrypts more of the handshake metadata. Future research will focus on maintaining accuracy under stricter encryption protocols and investigating whether flow association can compensate for the loss of handshake-level features.

Summary updated

Abstract

Many network services and tools (e.g. network monitors, malware-detection systems, routing and billing policy enforcement modules in ISPs) depend on identifying the type of traffic that passes through the network. With the widespread use of mobile devices, the vast diversity of mobile apps, and the massive adoption of encryption protocols (such as TLS), large-scale encrypted traffic classification becomes increasingly difficult. In this paper, we propose a deep learning model for mobile app identification that works even with encrypted traffic. The proposed model only needs the payload of the first few packets for classification, and, hence, it is suitable even for applications that rely on early prediction, such as routing and QoS provisioning. The deep model achieves between 84% to 98% accuracy for the identification of 80 popular apps. We also perform occlusion analysis to bring insight into what data is leaked from SSL/TLS protocol that allows accurate app identification. Moreover, our traffic analysis shows that many apps generate not only app-specific traffic, but also numerous ambiguous flows. Ambiguous flows are flows generated by common functionality modules, such as advertisement and traffic analytics. Because such flows are common among many different apps, identifying the source app that generates ambiguous flows is challenging. To address this challenge, we propose a CNN+LSTM model that uses adjacent flows to learn the order and pattern of multiple flows, to better identify the app that generates them. We show that such flow association considerably improves the accuracy, particularly for ambiguous flows. Furthermore, we show that our approach is robust to mixed traffic scenarios where some unrelated flows may appear in adjacent flows. To the best of our knowledge, this is the first work that identifies the source app for ambiguous flows.
BibTeX
@inproceedings{rezaei2019large,
  author = {Shahbaz Rezaei and Bryce Kroencke and Xin Liu},
  title = {{Large-scale Mobile App Identification Using Deep Learning}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2962018},
}
Figure