The threat of deepfakes is growing as AI tools to create them become more widely available. According to deepfake authentication provider Sumsub, there will be a 245% increase in deepfakes worldwide from 2023 to 2024, with the increase driven in part by the approaching election cycle. Deepfakes are also impacting the corporate sector, with a recent survey by Business.com finding that 10% of businesses have faced fraud involving deepfakes, such as cloned voices.
Not surprisingly, the trend has been a windfall for companies selling tools to protect against deepfakes and the technology used to create them. One such company, Pindrop, said on Wednesday it had secured a $100 million, five-year loan from Hercules Capital. CEO Vijay Balasubramaniyan said the money will be used for product development and hiring.
“Advances in generative AI, especially voice cloning, have made it a powerful tool,” Balasubramanian told TechCrunch. “Deepfake detection, powered by AI detection techniques, is becoming a necessity for all call centers to stay one step ahead of fraudsters.”
Pindrop develops anti-deepfake and multi-factor authentication products targeted at companies in banking, finance and related industries, and claims that its tools can identify callers in contact centers, for example, with better accuracy than competing solutions.
“Pindrop leverages a speech dataset of over 20 million utterances, including both real and synthetic speech, to train our AI models to distinguish between real human voices and synthetic speech,” Balasubramanian said. “We have also trained over 330 text-to-speech (TTS) models to identify TTS models used to create deepfakes.”
Bias is a common issue with deepfake detection models: Many audio models tend to recognize Western or American speech and perform poorly with different accents and dialects, which can lead detectors to classify legitimate audio as deepfake.
It's debatable whether synthetic training data (training data generated by the AI model itself) reduces or exacerbates bias; Balasubramaniyan believes it's the former, and he claims that Pindrop's voice authentication product focuses on the “acoustic and spectrotemporal features” of a voice, rather than pronunciation or language.
“AI-based speech recognition systems tend to display biased results for differences in tone, accent, dialect, and language that may have racial implications,” Balasubramanian said. “These biases stem from the homogeneity of the data used to train the systems, which may lack representation of different ethnicities, races, genders, and other differences, limiting the diversity of the data used to train AI systems.”
Regardless of the product's effectiveness, Pindrop has come a long way since ex-Googler Balasubramanian founded the company with former Barracuda Networks chief research officer Paul Judge and Mustaq Ahmad in 2011. Based in Atlanta and with about 250 employees, the company has raised $234.77 million in venture capital from investors including Vitruvian Partners, CapitalG, IVP and Andreessen Horowitz.
Asked why Pindrop chose debt over equity this time around, Balasubramanian said it was an “attractive option” to “efficiently fund growth” without diluting Pindrop's shares (a common strategy).
Balasubramanian added that proceeds from the loan will enable Pindrop to bring its technology to new sectors such as healthcare, retail, media and travel.
“With the advent of generative AI, we have seen a surge in demand for our solutions around the world and will be looking to expand into countries that face significant threats from deepfakes,” Balasubramanian said. “With our fraud prevention, authentication and biometric solutions, Pindrop is positioned to help businesses protect themselves and their consumers from the growing threat of fraud and deepfakes.”