Your 'Anonymized' Friend Network May Not Be as Private as Promised
Researchers built a lie detector for privacy claims — and some data-sharing tools fail it.
When companies and researchers want to share data about connections — your friend list, your bank transfers, your calls — they rarely hand over the raw files. Instead they use 'graph generators': programs that study the real network and spit out a fake one with the same overall shape. The pitch is that the copy keeps the useful patterns while hiding which specific people are linked. It's the same idea as publishing a realistic-looking map of a city rather than the actual home addresses.
The problem, according to this new paper, is that the promise is rarely checked properly. The team built an auditing method — think of it as a stress test or a lie detector — that tries to figure out whether two slightly different inputs produce tell-tale different outputs. If an attacker can tell them apart, information leaked. They ran three kinds of attacks: simple edge-guessing, attacks that look at local structure (who else is nearby), and attacks powered by GNNs, which are AI systems that learn from relationships rather than from isolated facts.
The AI-based attacks were the standouts. They picked up information that traditional statistics — simple counts like 'how many friends does this person have?' — completely overlooked. In other words, a network can pass the standard privacy checklist and still give away real secrets to a smart enough model. Leakage also varied by tool and by network, so there's no universal setting that makes everything safe.
What this means practically: privacy claims about connected data should be tested, not trusted. There's a catch — this is a lab study using two tools and two datasets, so it doesn't prove any specific company or app leaked your data. It simply gives regulators, auditors, and privacy teams a sharper tool for catching leaks before they reach the public.
- The study tests 'graph generators' — tools that fake a network so the real connections stay hidden — and finds some of them leak.
- AI models that learn from relationships spotted private links that traditional count-based privacy checks missed entirely.
- Leakage depended on both the tool and the network, meaning there's no one-size-fits-all privacy setting.
Why It Matters
Anonymized data about your contacts, calls, and payments may not be as anonymous as you were told.