Abstract: Big data and artificial intelligence have permeated discussions not only within business circles or academic spheres, but the considerations of people in all walks of life. Despite the vast discourse from talks, panels, journals, think pieces, and podcasts on these compelling new technologies, a crucial demographic had been surprisingly overlooked: children and adolescents. Decision-makers in business and government frequently display a lack of understanding and overlook the consequences of their actions on this vulnerable group, as evidenced by recent press releases, business memos, and legislation.
There is an urgent need to develop ethical policies that prioritize the protection of those providing data, especially minors who are particularly at risk. The dangers inherent in the handling of data and the use of AI are amplified when the humans involved are vulnerable, such as minorities, the poor, and crucially – the youth. If children are not being explicitly protected, then they are implicitly being left behind. This presentation will address the significant gap in research, legislation, and policy concerning the effects of data and AI on children and explore the necessity of safeguarding this critical population.
Abstract: Deepfakes—misleading content generated or manipulated using AI methods—have proliferated as vehicles for deception and fraud, posing ever-increasing threats to individuals and institutions. Audio deepfakes in particular are overlooked in existing literature compared to video and image counterparts (Khanjani et al., 2023). Our multidisciplinary team of data scientists and sociolinguists—experts in a subdiscipline of linguistics that deals with how human language varies, socially and stylistically—offers a novel approach to detecting audio deepfakes and other spoofed audio attacks by incorporating insights about spoken human language into machine learning techniques.
This talk shares results from four years of our ongoing research and outlines novel pathways for interdisciplinary collaboration to address deepfakes as a pressing societal problem. Findings demonstrate how audio representations, manually extracted by sociolinguists, increase the detection performance significantly at the scale of all types of spoofed audio attacks, when combined with machine learning models (Khanjani et al., 2023). Additionally, when AI models are used to automatically extract Audio Linguistic Representations designed for anti-Spoofing (ALiRaS), under expert supervision, the performance of the common baselines significantly increased. Overall, the talk demonstrates that leveraging human expert knowledge is crucial in creating robust audio representations used in spoofed audio detection for strengthening AI solutions.
