The Cyber Security Review | Tuesday, April 29, 2025
Fremont, CA: Artificial intelligence is sometimes viewed as a single domain encompassing a homogeneous set of solutions applicable to various kinds of application cases. However, the practical use of AI in real-world scenarios needs a level of specialization customized to certain fields of study, as each use case brings unique problems. In cybersecurity, applied AI has various specific challenges, and we will look at a few of the most pressing issues.
Lack of Labeled Data:
Stay ahead of the industry with exclusive feature stories on the top companies, expert insights and the latest news delivered straight to your inbox. Subscribe today.
In contrast to numerous other domains, the availability of data and labels in the cybersecurity sector could be improved, often necessitating highly specialized expertise for their generation. Examining a random assortment of logs from typical cybersecurity systems will likely yield no labels. Instances such as a user downloading a document are not classified as malicious or benign, and no data indicates whether a login attempt was legitimate. This situation is distinctive to the field of cybersecurity. In many other areas of applied artificial intelligence, labels are plentiful, facilitating techniques that capitalize on these labels.
Due to the scarcity of labels, most detection methodologies resort to unsupervised learning techniques, including clustering and anomaly detection, which rely on something other than labeled data. However, this approach presents significant drawbacks.
Anomalous is Not Malicious:
Numerous methodologies employ anomaly detection and clustering techniques to identify suspicious behaviors. Although these methods possess certain advantages, they also inadvertently identify numerous benign activities.
Examining any well-established network environment will reveal various assets and activities that are intentionally bizarre, such as vulnerability scanners, domain controllers, and service accounts, among others. These elements generate significant noise for anomaly detection systems and contribute to alert fatigue for Security Operations Center (SOC) analysts tasked with reviewing the alerts produced by these systems. Conversely, attackers often operate below the detection threshold, allowing them to evade notice, as the abnormal activity level required to fulfill their objectives is frequently much lower than that generated by the assets above.
Domain Expertise Is Critical and Hard to Find:
In contrast to numerous other fields, validating models in cybersecurity necessitates specialized knowledge. Determining whether a traffic light is green or red does not require an expert, but identifying whether a file is malicious demands the skills of a malware analysis professional. Developing AI models for cybersecurity relies on qualified experts who can validate outcomes and categorize instances to evaluate essential performance indicators (KPIs). Given the limited availability of such experts and the fact that supervised learning is the optimal approach for cybersecurity AI, this presents an additional significant challenge in effectively implementing AI within this domain.
More in News