Research pick: Do not feed the phish - "Prevention of cyber attacks and real-time social media spam detection and sentiment analysis using recurrent self-adaptive windowing approach"

Millions of us use social media and online messaging every day. It’s convenient but as with every tool, there are risks, such as the threat of crime such as spam and scam messages and phishing attacks that have the potential to cost the victims dearly. Malicious messages disrupt the user experience and pose serious security threats by attempting to steal personal data, distribute malware, and compromise digital systems. Standard detection tools often lag behind the continually evolving tactics employed by attackers, leaving users vulnerable to increasingly sophisticated scams.

Research in the International Journal of Information and Computer Security, introduces a new approach that could improve the detection of spam on digital platforms. Central to the research is the use of a recurrent neural network, which can analyse text, and not only detect patterns but recognise those same patterns when it subsequently encounters them.

The approach also embeds a so-called soft attention mechanism, which enables the neural network to prioritize parts of a message that are most relevant to identifying spam or malicious intent. This mechanism mimics the way humans naturally focus on keywords or suspicious phrases when quickly scanning content.

In addition, the system also uses self-adaptive windowing to detect content drift, where the scammers and spammers continually change the type of language they use to try to avoid detection. The system can through windowing update its learning progressively as new types of spam and scam messages emerge. This avoids the need for complete and frequent reboots, instead the system undergoes on-the-job training. This adaptability, the researchers suggest, is important for maintaining detection accuracy over time.

Tests showed the model could achieve 99.3 percent accuracy, which surpasses traditional methods such as decision trees and even naïve Bayes classifiers. The number of false negatives and false positives was negligible even when the system was tested on realistic, imbalanced datasets, where spam is far less common than legitimate messages, to ensure that it would work well in real-world conditions.

Patil, S.M., Mhatre, S., Dakhare, B. and Chavan, G.T. (2025) ‘Prevention of cyber attacks and real-time social media spam detection and sentiment analysis using recurrent self-adaptive windowing approach’, Int. J. Information and Computer Security, Vol. 27, No. 2, pp.261–284.