What Is Information Leakage?

Information leakage is the unintended disclosure of confidential information through the observable behavior or outputs of a system, as distinct from disclosure caused by an explicit breach of access control. A system may enforce every permission check correctly and still leak, because an adversary can infer protected values from timing, resource consumption, error messages, aggregate statistics, or the structure of the outputs the system was designed to release. Information security treats leakage as a property of the whole system rather than of any single component, since the leak often arises from an interaction the designers never modeled as a channel.

The concept was formalized beginning in the 1970s through work on covert channels in multilevel secure operating systems, and in 1982 through the noninterference definition, which holds that a system is secure when varying its high-classification inputs produces no change in its low-classification outputs. Strict noninterference proved too rigid for practical systems: a password checker that rejects a wrong guess necessarily reveals something about the stored secret, and a statistical database exists to reveal aggregate facts. That gap motivated a quantitative theory in which the question is how much information leaks rather than whether any does.

Quantitative Information Flow

Quantitative information flow models a system as a noisy channel from a secret input to an observable output and measures leakage as the reduction in the adversary's uncertainty about the secret. Shannon entropy gives one such measure, but min-entropy leakage and the more general g-leakage framework are usually preferred, because they express the adversary's chance of guessing the secret in one try and can be tuned to different operational threats. Leakage is then the difference between prior and posterior vulnerability, and a bounded leakage value can serve as a security requirement in the same way a bit-length serves as a cryptographic parameter. Game-theoretic extensions model defender and adversary as strategic players who each choose randomized policies, and work on information leakage games analyzes the resulting equilibria for systems where the defender can add noise at a cost.

Side Channels

Side channels are the dominant practical leakage vector. NIST characterizes a side-channel attack as one exploiting leakage from the physical implementation of a cryptosystem through timing, power consumption, or electromagnetic and acoustic emissions. Timing attacks recover keys from data-dependent execution paths; differential power analysis correlates power traces with hypothesized intermediate values; cache attacks such as Prime+Probe and Flush+Reload infer memory access patterns from shared microarchitectural state; and speculative execution attacks turn transient mispredicted execution into an observable cache footprint. Countermeasures include constant-time implementation, masking of intermediate values, partitioning of shared caches, and noise injection, and quantitative information flow analysis applied to hardware provides a way to compare their residual leakage rather than judging them qualitatively.

Leakage from Data and Models

Releasing aggregate data or trained models leaks information about the individual records behind them. Membership inference determines whether a specific record was in a training set, model inversion reconstructs attributes of training examples, and linkage attacks re-identify individuals in supposedly anonymized releases by joining quasi-identifiers against external data. Differential privacy addresses this class directly by bounding the influence any single record can have on an output distribution, expressed through the privacy loss parameter epsilon. Related concerns include verbose error messages, metadata retained in published documents, and traffic analysis that infers content from packet sizes and timing even under encryption.

Applications

Information leakage analysis has applications in a range of fields, including:

• Cryptographic hardware and smart card certification
• Processor microarchitecture design and security review
• Compiler and language-level information flow verification
• Privacy-preserving data publication and census release
• Machine learning privacy auditing
• Network traffic analysis and censorship resistance