Introduction

Your business could be impersonated in a phone call today. Deepfakes - digital reproductions of human voice and image via AI - are a scammer’s best asset, and it’s quickly becoming the most viable option in a fraud. There’s no coincidence that fraudulent calls, messages and emails have been getting more common. And it’s no wonder since the technology permits it. The best way to fight it is to position yourself as a leader in fraud detection & prevention.

Why it matters — now more than ever

Voice-based fraud is no longer an exception. According to a Truecaller report, impersonation scams involving cloned voices generate over $25 billion in losses annually. The technology required to replicate a voice is easily available online, either free or very low-cost, and can work with just a few seconds of recorded audio. Most systems still trust voice as a valid signal for identity.

In a high-profile case, Arup was scammed during a video conference where attackers used deepfake technology to replicate the CFO’s face and voice, leading to a $25 million fraud.

Certain demographics are more susceptible to fraudulent attacks. In a recent study by iProov, a staggering 0.1% of participants correctly identified all deepfake images and videos, even when they were specifically instructed to look for them. In real-world situations, recognition rates drop even further. 

Among people aged 65 and older, nearly 40% have never heard of deepfakes with factors like social isolation, cognitive decline, or lacking digital competencies leaving them especially vulnerable.

This is not a drill, it’s the reality we have at hand. Companies in Finance, Customer Support, Telecom and Healthcare industries are targeted the most. Imagine a scammer calling your call center with a cloned voice of a legitimate client. It’s not far-fetched. And when trust is lost, customers walk away.

In an environment where trust is everything, and reputations can be lost in an instant, real-time deepfake detection is no longer a luxury - it’s a critical line of defense.

AI can fight itself. Real time deepfake detection is possible!

A Romanian tech publication called Hyperplane wrote an article on how to build such a system where AI can detect other AI (deepfakes). It takes recordings of potentially fraudulent calls and returns a verdict: REAL or FAKE. 

The core of the deepfake voice detection system is a hybrid neural architecture that combines computer vision and sequence modeling. Incoming voice calls are converted into Mel spectrograms which are great for identifying audio artifacts left by synthetic generation methods. These spectrograms are then processed by a ResNet18 backbone used to detect spectral anomalies like flattened harmonics, unnatural pacing, and robotic textures that are hard for even high-quality generative models to fully mask.

To capture how these patterns evolve across time, like humans do, the system includes a Bi-directional GRU (Gated Recurrent Unit). This allows the system to distinguish organic variations in timing, emphasis, and prosody from the smoother, overly consistent patterns typical of synthetic speech.

Conclusions

It’s a matter of time until fraud finds you, if it hasn’t already. As AI advances, so does its dark side. Deepfake technology is no longer a futuristic threat, it’s a present-day weapon in the hands of cybercriminals. Real-time deepfake detection is a necessity for our days.

No system is completely immune, but with the right tools in place, the risk can be reduced. For a better understanding about the technical approach behind this solution, the full article on Hyperplane  provides a detailed look at the architecture and methodology used to detect deepfake audio in real time.

Article written by Silviu Gresoi - AI & Fraud Detection Specialist