Unmasking the Hidden Threats of Social Media: Detecting Misinformation, Bots, and Malicious Campaigns

The open nature of online social networks (OSNs) encourages communication, content creation, sharing, and community building. However, this openness, along with the ease of spreading unverified information, makes OSNs particularly susceptible to the spread of false information (or misinformation) by automated bot accounts through coordinated malicious campaigns. These dynamics combine to pose an increasing threat: social media manipulations (SMMs). Traditional techniques often address false information, bots, or campaigns individually, but each approach has significant drawbacks. Detecting false information is hindered by limited labeled datasets; social bot detection faces challenges from sophisticated bots that imitate human profiles; and campaign detection rarely differentiates between malicious and harmless bot activity over time. This thesis presents machine learning-based solutions to tackle these challenges across three key elements of SMMs. First, it introduces an automatic annotation model to reduce data scarcity by using fact-checked statements from PolitiFact, Snopes, and TruthOrFiction. By combining relevance and similarity voting methods, the model creates large-scale annotated datasets, which are then used to train stacked ensemble classifiers. When applied to 10.22 million COVID-19-related tweets, the approach generated an extensive ground-truth dataset, achieving detection performance of 83% precision, 96% recall, and a false positive rate of ≈4%; outperforming existing annotation and detection techniques. Second, BotSSCL, a self-supervised contrastive learning framework, is proposed to detect sophisticated social bots that closely resemble human users. By enhancing representation learning and improving linear separability in the embedding space, BotSSCL demonstrates ≈6% and ≈8% higher F1 scores than state-of-the-art methods on two tabular datasets, with robustness confirmed through sensitivity and ablation analyses. Finally, the thesis presents TBTrackerX, a framework for detecting trigger-based (TB) campaigns run by advanced bots (or agents) that activate only when specific keywords are present in user posts. Using a measurement-driven approach, the framework identifies malicious TB agents that exploit contextual replies, intermittent posting patterns, and dormancy after peak activity. The analysis also uncovers a coordinated TB ecosystem involving fake followers and shared masters. Collectively, these contributions advance the detection of false information, bots, and malicious campaigns, underscoring the urgent need for integrated moderation and defence mechanisms to safeguard OSNs against increasingly sophisticated SMMs.