[Paper Review] Privacy Failures in Encrypted Messaging Services: Apple iMessage and Beyond

Scott E. Coull, Kevin P. Dyer|arXiv (Cornell University)|Mar 8, 2014

Internet Traffic Analysis and Secure E-voting7 references3 citations

TL;DR

This paper demonstrates that end-to-end encrypted messaging services like Apple iMessage, WhatsApp, Viber, and Telegram leak significant user metadata—such as device OS, user actions, message language, and message length—through deterministic relationships between user interactions and encrypted packet sizes. Despite strong encryption, attackers achieve over 96% accuracy in inferring these details by analyzing packet lengths, revealing a critical privacy vulnerability across major messaging platforms.

ABSTRACT

Instant messaging services are quickly becoming the most dominant form of communication among consumers around the world. Apple iMessage, for example, handles over 2 billion message each day, while WhatsApp claims 16 billion messages from 400 million international users. To protect user privacy, these services typically implement end-to-end and transport layer encryption, which are meant to make eavesdropping infeasible even for the service providers themselves. In this paper, however, we show that it is possible for an eavesdropper to learn information about user actions, the language of messages, and even the length of those messages with greater than 96% accuracy despite the use of state-of-the-art encryption technologies simply by observing the sizes of encrypted packet. While our evaluation focuses on Apple iMessage, the attacks are completely generic and we show how they can be applied to many popular messaging services, including WhatsApp, Viber, and Telegram.

Motivation & Objective

• To investigate the extent of metadata leakage in end-to-end encrypted messaging services despite strong encryption.
• To determine whether packet size analysis can be used to infer user actions, message language, and message length across multiple platforms.
• To evaluate the generalizability of these attacks across popular messaging services like iMessage, WhatsApp, Viber, and Telegram.
• To assess the feasibility and cost of countermeasures such as random padding to mitigate these privacy leaks.
• To highlight the real-world privacy risks posed by metadata leakage in widely used messaging platforms.

Proposed method

• The researchers collected and analyzed network traffic from Apple iMessage, WhatsApp, Viber, and Telegram to observe encrypted packet sizes in response to user actions.
• They used machine learning techniques such as Naïve Bayes and linear regression to classify device OS, message language, and predict message length from packet sizes.
• A lookup table was constructed to map specific packet sizes to user actions like 'start typing' or 'send attachment'.
• The study evaluated the impact of random padding as a countermeasure by measuring the increase in packet size and its effect on attack accuracy.
• The researchers compared packet size distributions across services to identify deterministic patterns linking user behavior to encrypted traffic.
• They validated their findings using real-world data from iOS and macOS devices, analyzing over 100 packets per user session.

Experimental results

Research questions

• RQ1 Can an eavesdropper infer the device operating system of a messaging user by analyzing encrypted packet sizes?
• RQ2 To what extent can an attacker determine the specific user actions (e.g., typing, sending) based on packet size patterns?
• RQ3 How accurately can an attacker classify the language of a message using only encrypted packet sizes?
• RQ4 Is there a strong correlation between plaintext message length and the size of the encrypted packet across different messaging services?
• RQ5 What is the performance and cost of countermeasures like random padding in eliminating these privacy leaks?

Key findings

• An attacker can identify the device operating system (iOS or macOS) with 100% accuracy by analyzing encrypted packet sizes.
• User actions such as 'start typing' or 'send attachment' can be classified with 96% accuracy using a lookup table of packet sizes.
• The language of a message can be predicted with 98% accuracy using Naïve Bayes classification on packet size data.
• Message length can be predicted with a mean error of only 6.27 characters using linear regression on encrypted packet sizes.
• The same attacks are applicable to WhatsApp, Viber, and Telegram, with some services like Viber and WhatsApp showing even stronger leakage due to one-to-one packet size-to-message length mappings.
• Random padding can fully mitigate all attacks, but at a cost of over 300% increase in data overhead—adding at least a terabyte of extra traffic per day for services like iMessage.