An attack tailored to shared AI systems made backdoors more effective

The method used feedback from a shared model to create attack patterns that transferred more successfully from individual devices.

Editorial illustration — not from the study.

Federated learning trains an AI model across many devices or organizations while sharing model updates instead of collecting all the raw data in one place. The distributed setup can make it vulnerable to backdoor attacks, in which specially prepared inputs cause the trained model to produce an attacker-chosen result.

The proposed method, called Federated Generative Adversarial Trigger, uses publicly available versions of the shared model and a small number of samples to generate attack patterns. A fixed, trained shared model guides the pattern generator, helping the resulting triggers match features learned by the global model rather than only those of an individual device.

How the attack improved

The researchers report that their method reduced the loss that can occur when a backdoor created on an individual device is transferred into the global model. Across experiments on multiple datasets, the attack improved the attack success rate by 60.95–80.19% over the average performance of baseline attacks on the global model. The improvement was achieved without amplifying the malicious updates sent to the federated-learning system.

Why shared models are at risk

The findings show that attackers can use feedback from a shared model to tailor backdoor triggers to the features that model has learned. That makes the attack more effective during the transfer from local models to the global model, highlighting an additional security challenge for federated-learning systems used in settings such as connected devices.

Evidence and open questions

The evidence comes from experiments on multiple datasets comparing the proposed attack with baseline methods. The abstract does not identify the datasets, describe the tested defenses, give the number of experiments, or show whether the method has been tested in real-world federated-learning deployments. It therefore establishes the method’s reported performance in the evaluated experiments, but does not by itself show how it would perform against all systems or defenses.

High-mounted cameras spot pedestrians on the device at busy intersections

Researchers introduced a 13,314-image dataset for testing AI systems that view urban intersections from above. The results show that detection can run on edge devices such as the Jetson Orin Nano, although pedestrians remain harder to detect than vehicles and performance changes with scene conditions.

5 hours ago·2 views

AI & Computing

Where smaller AI models can work alongside—and compete with—larger ones

A survey examines the roles smaller AI models can play alongside large language models. It focuses on both collaboration and competition, including how smaller systems may help researchers and businesses use computing resources more efficiently.

22 hours ago·6 views

AI & Computing

AI watermarks in Europe do not yet meet all four legal tests

The European Union’s AI Act calls for providers to mark and detect outputs from general-purpose AI models. A new analysis finds that no current watermarking approach for large language models meets all four of the law’s stated standards: reliability, interoperability, effectiveness and robustness.

2 days ago·17 views