Skip to main content

Attribution

Attribution is the process of linking an action to an actor. For a privacy-focused operator, attribution is the adversary's goal: they do not need to stop the action, they need to prove who did it. This file covers how attribution works, what features adversaries use, and how to make attribution harder.

This file is conceptual. It does not cover tool installation or configuration. Those live in the hands-on subcategories (01-06). This file covers the reasoning behind attribution resistance.

1. Overview​

Attribution answers the question "who did this?". It is distinct from deanonymization: deanonymization reveals an identity, attribution links that identity to an action. An adversary can attribute an action without ever knowing the actor's real name - a persistent pseudonym is enough for legal, financial, or social consequences.

Attribution is probabilistic, not binary. Adversaries build a case from multiple independent features. A single feature is rarely enough; a pattern of features is. Attribution resistance is therefore the practice of minimizing the number of independent features an adversary can collect.

2. Threat Model​

The adversary for attribution is an investigator with time and resources. This adversary is not necessarily a nation-state; it can be a platform trust-and-safety team, a civil litigant, a journalist, or a competitor. The features they collect fall into six categories:

  • Technical - IP addresses, device fingerprints, browser fingerprints, TLS fingerprints, timing patterns.
  • Linguistic - vocabulary, syntax, punctuation, typos, code style, formatting habits.
  • Behavioral - activity hours, response latency, order of operations, tool choices.
  • Social - relationships, references, shared interests, common contacts.
  • Financial - payment methods, crypto addresses, purchase patterns.
  • Physical - location, movement, access to specific infrastructure.

The adversary does not need all six. Two or three independent features are often enough to build a case that survives legal or public scrutiny. The threat model for attribution resistance is therefore: assume the adversary collects all six, and design so that no two can be linked.

3. How It Works (technical)​

Attribution works by correlation across independent data sources. The adversary collects features from each source, then looks for intersections. The sources are:

  • Network logs - ISP records, VPN provider records, Tor relay logs, CDN logs, platform access logs.
  • Platform data - account metadata, login history, device identifiers, content history.
  • Public data - forum posts, social media, code repositories, domain registrations, certificate transparency logs.
  • Financial data - bank records, exchange records, blockchain analysis.
  • Physical data - CCTV, badge logs, travel records, purchase records.
  • Human data - informants, colleagues, family, social engineering.

The adversary's method is Bayesian: each feature updates a probability. A feature that is rare in the population is more informative than a feature that is common. The adversary's goal is to reach a probability high enough to act (legally, financially, or socially).

Attribution resistance is therefore not about being invisible. It is about being one of many plausible actors, so that the adversary's probability never reaches the action threshold.

4. Practical Framework​

A working attribution-resistance framework, in six steps:

  1. Identify the features. For each action, list the features the adversary could collect. Be exhaustive: technical, linguistic, behavioral, social, financial, physical.

  2. Classify by rarity. For each feature, estimate how rare it is in the relevant population. A rare feature is a strong signal; a common feature is weak. Prioritize the rare features.

  3. Separate the features. For each rare feature, decide how to separate it from the action. This is the compartmentalization policy (see 02 Compartmentalization).

  4. Introduce noise. Where separation is not possible, introduce noise: vary timing, vary phrasing, vary tools, vary routines. Noise reduces the signal-to-noise ratio the adversary sees.

  5. Test the case. Assume the adversary has collected every feature you could not separate or noise. Build the strongest case they could build. If the case is strong, the resistance is insufficient.

  6. Re-evaluate. Attribution techniques evolve. New correlation methods (machine learning on writing style, blockchain clustering) appear regularly. Re-evaluate on a schedule.

The hardest part is honesty: the operator must assume the adversary is competent and patient, not lucky.

5. Real-World Examples​

The Unabomber case (1995). Ted Kaczynski was attributed through linguistic analysis: his brother recognized the writing style in the published manifesto. The technical compartment was strong; the linguistic compartment was not.

The "BTK" case (2005). Dennis Rader was attributed through a floppy disk he sent to the police. The disk metadata linked the file to a computer at his church, which linked to him. The physical compartment was strong; the metadata compartment was not.

The "Golden State Killer" case (2018). Joseph DeAngelo was attributed through genetic genealogy: a distant relative's DNA upload to a genealogy site allowed investigators to build a family tree and narrow it to him. The physical compartment was strong; the genetic compartment was not.

The "Silk Road" case (2013). Ross Ulbricht was attributed through a combination of a personal email in an early forum post, a Stack Overflow question, and a login from a non-anonymized IP. The technical compartment was strong; the historical and behavioral compartments were not.

The "LulzSec" cases (2011-2013). Multiple members were attributed through IRC logging, informants, and a single login from a non-anonymized IP. The technical compartment was strong; the social and endpoint compartments were not.

The "Ashley Madison" investigation (2015). The breach did not attribute the users; it revealed them. But the subsequent investigations attributed specific users through payment records and email addresses. The platform compartment was strong; the financial and identity compartments were not.

6. Common Mistakes​

  • Assuming technical anonymity is enough. Attribution is multi-feature. Technical anonymity protects one feature, not all.
  • Ignoring linguistic features. Writing style is a fingerprint. Machine-learning stylometry can link texts that share no technical infrastructure.
  • Ignoring behavioral features. Activity hours, response latency, and order of operations are fingerprints. They persist across technical changes.
  • Ignoring social features. The people you talk to, the topics you discuss, and the references you make are fingerprints. They persist across identity changes.
  • Ignoring financial features. Payment methods and crypto addresses are fingerprints. Blockchain analysis is increasingly effective.
  • Underestimating the adversary's patience. Attribution cases are built over years. A single slip in year five can link everything from year one.
  • Assuming that "they cannot prove it" is protection. Attribution does not require proof in a court. It requires enough probability for the adversary to act: a platform ban, a civil suit, a public accusation, a financial restriction.

7. Attacks and Detection​

An adversary attacks attribution resistance by collecting more features from more sources. The attacks are:

  • Stylometry. Machine-learning analysis of writing style. Links texts by vocabulary, syntax, punctuation, and formatting.
  • Timing analysis. Correlation of activity timestamps across sources. Links identities by shared activity windows.
  • Fingerprinting. Browser, device, and TLS fingerprinting. Links sessions by shared technical features.
  • Blockchain analysis. Clustering of crypto addresses by shared transaction patterns. Links wallets by shared behavior.
  • Social network analysis. Mapping of relationships and references. Links identities by shared contacts and interests.
  • Human intelligence. Informants, social engineering, and physical surveillance. Links identities through people.

Detection of attribution risk is proactive: the operator should assume the adversary is collecting features continuously, and periodically audit which features have been exposed. The defensive move is to minimize the number of exposed features, not to hide the exposed ones.

8. Risk Context​

Attribution resistance has a cost: more compartmentalization, more noise, more discipline. The cost must be proportional to the adversary. For a low-capability adversary (platform trust-and-safety, scraper), basic hygiene is enough. For a medium-capability adversary (civil litigant, investigative journalist), stronger separation is required. For a high-capability adversary (law enforcement with warrant, national CERT), the cost is high, and perfect resistance may be unachievable through consumer tools alone.

The risk of under-resisting is exposure. The risk of over-resisting is operational error: too much noise makes the operator's own work inconsistent, and inconsistency is itself a signal. The correct balance is the one the operator can maintain without breaking their own routines.

9. Verification​

Attribution resistance is verified by adversarial reconstruction:

  • Stylometry test. Compare the writing style of the protected identity to the writing style of any other identity the operator maintains. If a machine-learning model can link them, the resistance is insufficient.
  • Timing test. Compare activity windows across identities. If they overlap, the resistance is insufficient.
  • Fingerprint test. Compare browser, device, and TLS fingerprints across sessions. If they match, the resistance is insufficient.
  • Blockchain test. Compare crypto addresses and transaction patterns. If they cluster, the resistance is insufficient.
  • Social test. Compare contacts, references, and interests. If they overlap, the resistance is insufficient.
  • Human test. Assume the adversary has an informant in the operator's social circle. What would the informant reveal? If the answer links identities, the resistance is insufficient.

If any test succeeds, the resistance is incomplete. Fix the exposed feature, then re-test.

Attribution resistance is a defensive discipline. It is used by journalists protecting sources, activists protecting themselves, researchers protecting participants, and ordinary people protecting their privacy. It is not illegal.

The legal risk is in the actions the resistance enables. If the resistance is used to evade legitimate law enforcement, the resistance itself may be used as evidence of intent. If it is used to protect lawful privacy, it is protected speech in most jurisdictions.

Operators in jurisdictions with restrictive speech or privacy laws should consult a lawyer before documenting attribution-resistance policies that could be seized. The policy is a document; the document can be evidence.

CAUTION

This material is for educational purposes and for authorized, lawful privacy protection only. It does not constitute legal advice. Laws differ by jurisdiction. You are responsible for understanding and following the laws that apply to you.

11. References​

  • The Tor Project, "Tor: Overview" and "Tor: Metadata" documentation.
  • Electronic Frontier Foundation, "Surveillance Self-Defense" guide.
  • Citizen Lab, "The Not-So-Secret Life of Anonymity" (2014).
  • Arvind Narayanan et al., "On the Feasibility of Internet-Scale Author Identification" (2008).
  • Sadia Afroz et al., "Detecting Hoaxes, Frauds, and Deception in Writing Style" (2012).
  • Sarah Zhang, "The Golden State Killer Case and the Future of Genetic Privacy" (The Atlantic, 2018).
  • The Intercept, "The Silk Road Investigation" (2015).
  • United States v. Ulbricht, trial record (2015).
  • The Guardian, "LulzSec: The Inside Story" (2013).