Australian researchers unveil tool to detect audio deepfakes
Researchers from Australia’s national science agency have developed a technique that could help counter the surge in voice-cloning fraud, which cost victims more than $410 million in the first half of 2025.
The method, called Rehearsal with Auxiliary-Informed Sampling, or RAIS, was announced Tuesday by scientists from CSIRO, Federation University Australia and RMIT University. The technology determines whether an audio clip is genuine or artificially generated while maintaining accuracy as new attack methods emerge—a challenge that has plagued existing detection systems.
Addressing Catastrophic Forgetting
Unlike traditional detection systems that must be retrained from scratch when new deepfake techniques emerge, RAIS employs what researchers call “continual learning.” The system automatically stores a diverse set of past examples, including audio traits imperceptible to humans, enabling it to recognize new deepfake styles without losing knowledge of older ones.
”If you just fine-tune on the new samples, it will cause the model to forget the older deepfakes it knew before,” said Dr. Kristen Moore from CSIRO’s Data61, one of the study’s authors. The technique uses auxiliary labels—markers beyond simple “fake” or “real” classifications—to maintain a rich mix of training data.
In testing, RAIS achieved an average error rate of 1.95 percent across five consecutive learning experiences, outperforming competing methods. The code is now available on GitHub.
Rising Threat Landscape
Audio deepfakes have emerged as a tool for bypassing voice-based biometric authentication, impersonation and disinformation campaigns. In February, criminals used an AI-cloned voice of Italy’s Defense Minister Guido Crosetto to convince a former Inter Milan owner to transfer nearly one million euros for what was falsely presented as a government operation to free hostages.
Cybersecurity experts also believe AI voice cloning may have enabled a June breach of Qantas systems that exposed data from nearly six million customers. The suspected attackers, a group called Scattered Spider, have a documented history of using synthetic voices to trick help desk staff into handing over credentials.
”Audio deepfakes are evolving rapidly, and traditional detection methods can’t keep up,” said Falih Gozi Febrinanto, a recent PhD graduate of Federation University Australia who co-authored the research. “RAIS helps the model retain what it has learned and adapt to new attacks.” (perplexity.ai)
