Securing the Unseen: Can AI Catch Cyberattacks Nobody Has Seen Before?

By Nuruzzaman Faruqui

·

A zero-day attack exploits a weakness nobody knew existed, so there is no signature or rule to catch it. This review asks which kinds of AI are best at spotting these unknown attacks, and how fairly they are being tested.

4AI model families compared
99.56%best reported accuracy (DNNs, CICDDoS2019)
PRISMAsystematic review protocol

The problem

Traditional security tools recognise known threats by their signatures. Zero-day exploits appear in unknown forms, so they slip past. Many AI studies report very high accuracy, but often on tests that do not reflect truly unseen attacks.

What the research did

We carried out a systematic review of four families of AI models for zero-day intrusion detection: convolutional neural networks, deep neural networks, Bayesian networks and reinforcement learning. We followed a PRISMA-style protocol and paid special attention to “zero-day faithful” tests, such as splitting data by time or testing on a different dataset.

Beyond accuracy, we looked at practical measures: detection rate at a fixed false-alarm rate, how well-calibrated the probabilities are, time to alert, throughput and memory footprint.

What we found

Each family has its strengths. Deep neural networks led on feature-rich inputs (about 99.56% reported accuracy on CICDDoS2019). CNNs were efficient at the network edge (92.17% on Bot-IoT). Bayesian networks offered interpretable uncertainty (99.74% on NSL-KDD). Reinforcement learning showed promise for adaptive detection and response (96.18% on CSE-CIC-IDS2018).

We proposed a consistent, leakage-aware way to evaluate models so future studies can be compared fairly, plus practical advice on fitting models to data, choosing operating points and meeting latency budgets.

Who benefits

  • Security teams choosing AI for unknown-threat detection
  • Researchers, who get a fairer evaluation framework
  • Organisations facing advanced, targeted attacks

Limits and what comes next

Figures come from different studies with different datasets, so they are not head-to-head comparisons. That is exactly why the review calls for a common evaluation protocol.

The team

This review was led by Abdullah Al Siam (Daffodil International University) with me, AKM Azad (Imam Mohammad Ibn Saud Islamic University) and Mohammad Ali Moni (Charles Sturt University).

Frequently asked questions

What is a zero-day attack?

An attack that exploits a software weakness unknown to the vendor, so no fix or signature exists yet.

Can AI detect zero-day attacks?

AI can spot unusual behaviour that may indicate unknown attacks, though performance depends heavily on how it is tested.

Which AI is best?

It depends on the setting: deep networks for rich data, CNNs for edge devices, Bayesian networks for interpretability.

About the author

Nuruzzaman Faruqui is an award-winning AI educator and researcher who has taught artificial intelligence and machine learning at university level since 2018. A two-time winner of the Excellence in Teaching Award (2023, 2024) and winner of the Best Researcher Award (2024) at Daffodil International University, he is now a doctoral researcher in AI at Monash University, Australia. He has published 53 papers and six books, and writes these articles to make AI research easy for everyone to understand. His teaching · About Zaman · Get in touch

Keep reading