Phishing detectors can leak webpage details through shared data

A new attack inferred private webpage information from the intermediate data sent to servers by some language-based phishing detection systems.

Editorial illustration — not from the study.

Split learning is designed to keep raw webpage data on users’ devices while sending only intermediate information to a server. Researchers report that this shared information can still reveal private webpage content in language-based phishing detection systems.

Their Semantic Information Reconstruction Attack uses large language models to infer sensitive webpage details from this shared data. In tests on real-world phishing datasets, it outperformed conventional reconstruction attacks, pointing to a privacy weakness that developers may need to address.

What the attack recovered

The paper introduces the Semantic Information Reconstruction Attack, or SIRA, to infer sensitive semantic elements—meaning private details and themes—from the intermediate data shared by split-learning language models. In experiments using real-world phishing datasets, SIRA more accurately inferred private webpage information than conventional reconstruction attacks. The results indicate that keeping raw inputs on client devices does not by itself guarantee privacy in these systems.

Why the privacy risk matters

Split learning is used to reduce the computing burden on devices while avoiding direct transmission of raw inputs. The study shows that the intermediate data sent to the server may still disclose information about webpages used by phishing detectors, creating a privacy concern for security applications. The findings support developing defenses specifically for language-based split-learning systems.

Evidence and caveats

The findings come from experiments using real-world phishing datasets, in which the new attack was compared with conventional reconstruction attacks. The abstract does not report the datasets’ sizes, the exact accuracy results, or how the attack performs against specific privacy protections. It also describes a potential vulnerability in language-model-based split-learning systems, rather than evidence that every such system leaks the same information.

Reliable AI partnerships depend on people, not just stronger models

A survey of human-AI collaboration argues that more capable AI systems do not automatically produce reliable partnerships. It identifies human-centered design, governance and attention to high-stakes uses as central to making these systems trustworthy and beneficial.

9 hours ago·6 views

AI & Computing

RNA splicing tools can follow irrelevant clues and miss RNA’s shape

Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.

20 hours ago·5 views

AI & Computing

Better-trained AI finds more flaws in unfamiliar software projects

AI vulnerability detectors often perform less well on software projects outside their training data. A study finds that cleaner, more varied training data and an encoder-based design improved detection, including a reported 6.8% gain in recall on the BigVul benchmark.

2 days ago·33 views