Q: What ethical considerations justify hiding the sender’s identity in audio deepfakes, especially when the technology is used to create groundbreaking content?
AND: Exploring why research plays an vital role in concealing the identity of the source broadcaster, despite, for example, the widespread exploit of generative models to create sound in entertainment, raises ethical considerations. Speech is not just about “who are you?” (identity) or “what are you saying?” (contents); It contains tons of sensitive information, including age, gender, accent, current health, and even hints about upcoming future medical conditions. For example, our latest research article on “Detection of dementia based on long neuropsychological interviews” shows the possibility of detecting dementia based on speech with significantly high accuracy. Moreover, there are many models that can detect gender, accent, age and other information from speech with very high accuracy. There is a need for advances in technology to prevent the unintentional disclosure of such private data. Trying to anonymize the identity of a source speaker is not only a technical challenge, but also a moral obligation to protect individual privacy in the digital age.
Q: How can we effectively address the challenges posed by audio deepfakes in spear-phishing attacks, given the risks involved, developments in countermeasures, and advances in detection techniques?
AND: The exploit of audio deepfakes in spear-phishing attacks poses many risks, including the spread of disinformation and imitation news, identity theft, privacy breaches, and malicious content alteration. The recent proliferation of duplicitous robocalls in Massachusetts is an example of the harmful impact of such technology. We also recently talked to he talked with about this technology and how easily and inexpensively such false sounds can be generated.
Anyone without a mighty technical background can easily generate this sound using many tools available online. Such imitation news from deepfake generators can distort financial markets and even election results. Voice theft to gain access to voice-controlled bank accounts and the unauthorized exploit of voice identities for financial gain serve as a reminder of the urgent need for decisive countermeasures. Further risks may include privacy breaches, where an attacker could exploit the victim’s audio without their consent. Moreover, attackers can also change the content of the original audio file, which can have earnest consequences.
Two main and dominant directions have emerged in the design of false sound detection systems: artifact detection and liveness detection. When sound is generated by a generative model, the model introduces some artifacts into the generated signal. Scientists design algorithms/models to detect these artifacts. However, this approach presents some challenges due to the increasing complexity of audio deepfake generators. In the future, we may also see models with very little or almost no artifacts. Liveliness detection, on the other hand, exploits inherent features of natural speech, such as breathing patterns, intonations, and rhythms, which are complex to accurately reproduce in AI models. Some companies, such as Pindrop, are developing such solutions to detect audio fakes.
Additionally, strategies such as audio watermarking serve as proactive protection by embedding encrypted identifiers in the original audio to trace its origins and prevent tampering. Despite other potential vulnerabilities, such as the risk of replay attacks, ongoing research and development in this area offers promising solutions to mitigate the threats posed by audio deepfakes.
Q: What are the positive aspects and benefits of deepfake audio technology, despite their potential for misuse? How do you envision the future relationship between artificial intelligence and our experiences of sound perception evolving?
AND: Unlike the primary focus on the nefarious uses of deepfake audio, the technology holds enormous potential to positively impact a variety of sectors. Beyond the realm of creativity, where voice conversion technologies provide unprecedented flexibility in entertainment and media, audio deepfakes could bring change to the healthcare and education sectors. For example, my current work on anonymizing patient and physician voices in cognitive interviews in health care makes it easier to share key medical data for research around the world while ensuring privacy. Sharing this data among researchers promotes developments in the areas of cognitive health care. The exploit of this technology in voice restoration offers hope for people with speech impairments, such as ALS or dysarthric speech, by improving communication capabilities and quality of life.
I am very positive about the future impact of sound-generating AI models. The future interplay of artificial intelligence and sound perception is poised for groundbreaking advances, particularly from the perspective of psychoacoustics – the study of how humans perceive sounds. Innovations in augmented and virtual reality, exemplified by devices like Apple Vision Pro and others, are pushing the boundaries of the audio experience towards unparalleled realism. Recently, we have seen an exponential raise in the number of sophisticated models appearing almost every month. This rapid pace of research and development in this field holds the promise not only of improving these technologies, but also of expanding their applications in ways that will have profound benefits for society. Despite the inherent risks, the potential for audio-generating AI models to revolutionize healthcare, entertainment, education, and more demonstrates the positive trajectory of this field of research.