Can AI listen to you? It seems like. Researchers from renowned universities in the UK have developed a groundbreaking AI system that can eavesdrop on keyboards to collect potentially sensitive data. This algorithm, described in a recent article, achieved an impressive 93-95% accuracy in detecting key presses based solely on audio recordings from a MacBook Pro keyboard.
TL;DR:
- AI system that can eavesdrop on keyboards and detect pressed keys with an accuracy of 93-95% based on audio recordings from a MacBook Pro keyboard.
- Microphones in consumer devices enable acoustic side channel attacks, posing a risk to data security.
- Protect against AI eavesdropping by confusing the model with different typing styles or using touch screen keyboards. Stay vigilant for emerging AI threats to protect data and privacy.
The AI Listening Study
The study sheds light on the widespread use of microphones in consumer devices that can be misused for acoustic side channel attacks. The consumer devices are phones, laptops, tablets, etc. Therefore, the potential exploitation poses a risk to data security. Unlike previous approaches to laptop keystroke detection via audio, this AI-based method has incredible precision.
According to the researchers, their AI model is much more accurate than other hardware-based methods. Other methods often face limitations due to distance and bandwidth limitations. However, with microphones embedded in common consumer gadgets, the acoustics of typing have become increasingly fragile and accessible.
So how does this advanced audio algorithm work? The researchers recorded audio samples of typing on a MacBook Pro. They pressed each key 25 times. This allowed the AI to distinguish slight variations in the sound emitted by each key.
Then the audio recordings were converted into spectrograms. Spectrograms provide visual representations of sound frequencies over time. In addition, the AI was trained on these spectrograms, learning to correlate different patterns with different keystrokes.
So applying this training to thousands of segments of audio gives the algorithm an incredible understanding of the acoustic fingerprints unique to each key. Then, once trained on a specific keyboard model, the AI can accurately analyze new audio recordings. The recordings allow it to predict keystrokes – and know what you’re typing. The AI with MacBook Pro keyboard training achieved an impressive 93-95% precision rate. Even when the noise came from Zoom call recordings, the AI’s performance dropped only slightly.
How can I protect myself against AI?
There is no reason to panic (yet). The AI system must be calibrated for specific keyboard models and audio environments. So if a hacker wants to use AI to collect data, they must have access to the necessary training data. But if you want to stay extra safe, there are a few actions you can take. First, confuse the AI model by pressing keys differently. This reduces the accuracy by 40%. Second, use speakers to play sounds. You can even swap your keyboard for a touch screen keyboard.
As the capabilities of AI grow, so does concern about emerging threats. That’s why it’s important to keep up to date with the latest AI developments and to protect your data and privacy.