As people increasingly leverage AI tools like ChatGPT to improve their productivity, cybersecurity experts have warned that malicious actors are also following this trend. A study by UK-based researchers has demonstrated how a machine learning algorithm can determine what a person is typing by simply analyzing keystroke audio.
The study, published on August 3, involved training a state-of-the-art AI model on the keystroke sounds of a laptop recorded on a nearby smartphone mic.
The researchers said cybercriminals and hackers could use a similar AI tool to steal passwords and other sensitive information with just an audio recording.
While they have not observed this threat in the wild yet, the ubiquity of smartphone microphones and the success of the model raises concerns.
Model Predicts Text With 95% Accuracy
The three researchers from universities in the UK essentially used an AI model to carry out an acoustic side-channel attack. The attack involved eavesdropping on a victim’s acoustic signals and stealing information.
The researchers trained an AI model on the laptop keystrokes picked up by a phone microphone and over Zoom. Worryingly, they achieved a great degree of success. The model achieved an accuracy of 95% with a phone microphone and 93% accuracy over Zoom.
“Our results prove the practicality of these side channel attacks via off-the-shelf equipment and algorithms,” the paper states.
The researchers ran their experiments on a MacBook Pro 16-inch (2021) with 16GB of memory and an Apple M1 Pro processor. They used an iPhone 13 mini as the phone recording device. The idea was to use devices that are easily available to highlight the ease with which a hacker can carry out a similar attack.
Possible mitigation techniques
They also provided mitigation techniques to protect users from such attacks. The use of randomized passwords, or strong passwords, along with a different typing style, could prevent the algorithm from accurately picking up the password, the researchers said.
“The second simple defence against such attacks would be the use of randomised passwords featuring multiple cases. With the success of language-based models in [mentioned examples], passwords containing full words may be at greater risk of attack,” the paper reads.
“When touch typing was used, [mention experiment] saw keystroke recognition reduce from 64% to 40%, which (while still an impressive feat) may not be a high enough accuracy to account for a complex input featuring the shift key, backspace and other non-alphanumeric keys.”
“Additionally, a change in typing style may be implemented alongside mitigation techniques presented in other papers and requires no software or hardware component,” the researchers stated.
They also recommend that Voice-over-IP (VoIP) services work on automatic suppression or removal of keystroke acoustics from their applications.
AI Won’t Replace Hackers Just Yet, Says NordVPN CTO
Meanwhile, experts at NordVPN have warned that AI has enabled hackers to improve and increase their malicious activity. The number of cyberattacks and their sophistication has doubled since the introduction of ChatGPT. Criminals rely on AI tools to enhance spear-phishing attacks and modify their malware in real time.
“AI will not steal jobs from hackers, at least soon,” said Marijus Briedis, CTO at NordVPN.
“Cybercriminals are keen users of AI-driven tools, but it’s about improvement, not replacement. Hackers learned how to use AI to increase the capacity of their work and make their job easier, quicker, and more effective.”
“The utilization of AI tools has facilitated the automation of a significant portion of phishing attacks, and it is anticipated that the frequency of such attacks will escalate in the future, posing a significant cybersecurity threat,” Briedis added.
Nord recommends protecting yourself with antivirus and VPN
Since AI tools can digest vast amounts of data, hackers can rely on them to create individually tailored phishing attacks in very short periods. The texts of these phishing communications can be very convincing, making it even more important for users to check URLs and ascertain the sender’s identity before clicking a link or responding to an email.
Furthermore, attackers can use AI algorithms to modify their malware in real-time to bypass antivirus software and other defense mechanisms.
“With this kind of automation, hackers are seriously challenging traditional cybersecurity tools and exploiting their vulnerabilities,” Briedis said.
NordVPN recommends using a reliable antivirus with advanced protection against modern threats and also enabling a firewall. Furthermore, users must protect their privacy with a VPN when using public WiFi.
