Beware the Deepfake Threat: Study Reveals Alarming Inability to Detect Synthetic Speech
Criminals have the ability to deceive individuals using deepfake speech technology, as revealed by a new study conducted in the United Kingdom. The research highlights how people struggle to detect more than one in four deepfake speech samples, opening the door for potential exploitation. This study stands out as the first to assess human capability in recognizing artificially generated speech in languages other than English.
Deepfakes are a form of synthetic media designed to imitate the voice or appearance of a real person. Criminals have misused this technology to defraud unsuspecting individuals of large sums of money. These deepfakes are a product of generative artificial intelligence (AI), a branch of machine learning (ML) that trains algorithms to identify patterns and characteristics from datasets such as video or audio recordings of real people. This training allows for the generation of authentic-sounding audio or visual content.
While earlier versions of deepfake speech algorithms required thousands of voice samples to generate original audio, the most recent pre-trained algorithms can replicate a person’s voice using just a three-second clip of them speaking. Tech giant Apple even released software for iPhones and iPads that enables users to create copies of their voices using only 15 minutes of recordings.
To determine the human ability to differentiate between real and deepfake speech, researchers at University College London (UCL) employed a text-to-speech (TTS) algorithm trained on publicly available datasets in English and Mandarin. Using this algorithm, they created 50 deepfake speech samples in each language that were distinct from the training samples.
A total of 529 participants were presented with both real and deepfake speech samples to test their ability to distinguish between them. Shockingly, participants only managed to identify deepfake speech with an accuracy of 73 percent, as reported in the journal PLoS ONE.
The research team observed only a slight improvement in detection rates even after training the participants to recognize aspects of deepfake speech. Interestingly, English and Mandarin speakers demonstrated similar rates of detection, although they mentioned different speech features as indicators of phoniness. English speakers frequently mentioned aspects such as breathing, while Mandarin speakers referred to cadence, pacing, and fluency.
Kimberly Mai, the study’s first author from UCL Computer Science, warns, Our findings confirm that humans are unable to reliably detect deepfake speech, whether or not they have received training to help them spot artificial content. It’s also worth noting that the samples that we used in this study were created with algorithms that are relatively old, which raises the question whether humans would be less able to detect deepfake speech created using the most sophisticated technology available now and in the future.
Considering the potential harm deepfake technology can cause to individuals and societies, there is a growing concern that criminals and even nation states may exploit this tool. Instances of criminals employing deepfake speech have already been documented, such as a case in 2019 where the CEO of a British energy company transferred a substantial amount of money to a fraudulent supplier after being tricked by a deepfake recording of his boss’s voice.
While there are undeniably benefits to generative AI audio technology, including enhanced accessibility for individuals with speech impairments or those who may lose their voice due to illness, we must be cautious about the risks. Governments and organizations must develop strategies to address the potential abuse of these tools. At the same time, we shouldn’t overlook the positive possibilities that lie ahead, emphasizes Professor Lewis Griffin, the senior author of the study.
The research team is currently working on developing more effective automated speech detectors to combat the deepfake threat.
As deepfake technology continues to advance and becomes more easily accessible, it is crucial for individuals, governments, and organizations to stay vigilant and informed about this emerging threat. The ability to detect deepfake speech poses a significant challenge that needs to be urgently addressed to safeguard individuals and maintain technological integrity.
