Redazione RHC : 20 October 2025 08:18
A team of researchers in Hong Kong has released a system called Lancelot, which represents the first practical implementation of federated learning while also being protected from data tampering attacks and confidentiality breaches.
Federated learning allows multiple participants (clients) to jointly train a model without revealing the source data. This approach is particularly important in medicine and finance, where personal information is strictly regulated.
However, these systems are vulnerable to data poisoning : an attacker can upload fake updates and distort the results . Federated learning methods have partially solved this problem by discarding suspicious updates, but they have not protected against the possible recovery of encrypted data from the neural network’s memory.
The team decided to combine cryptographic security and attack resistance. Lancelot uses fully homomorphic encryption to ensure that all local model updates remain end-to-end encrypted.
The system also selects trusted client updates without revealing who is trusted. This is achieved using a special “masked sorting” mechanism : a trusted key center receives the encrypted data, sorts the clients by trust level, and returns only an encrypted list to the server, obscuring the training participants. This way, the server aggregates only verified data without revealing its origin.
To speed up the calculations, the developers implemented two optimization techniques. Lazy relinearization postpones costly cryptographic steps until the final stage, reducing the CPU load. Dynamic hoisting groups repetitive operations and executes them in parallel, even on GPUs, significantly reducing overall training time.
The result is a solution that addresses two vulnerabilities of federated learning: it is resilient to malicious attacks while ensuring complete data confidentiality . Tests have shown that Lancelot not only prevents data leaks and sabotage, but also significantly reduces model training times by optimizing cryptographic operations and leveraging GPUs.
The researchers intend to expand the Lancelot architecture, making it suitable for large-scale scenarios. Potential applications include training AI systems in hospitals, banks, and other organizations that handle sensitive data. The team is currently testing new versions with support for distributed keys (threshold and multi-key CKKS), the integration of differential privacy methods, and asynchronous aggregation, which will allow the system to operate reliably even with unstable network connections and a wide variety of client devices.
Redazione