Adaptive and trustworthy knowledge discovery is critical in complex, real-world social media environments, where cyberbullying behaviors frequently evolve in covert and diverse forms. Traditional detection methods, often limited to single-modal analysis, struggle to reliably identify implicit bullying cues embedded within multimodal user data, highlighting the necessity for adaptive and privacy-preserving solutions. In this paper, we present a federated, multimodal cyberbullying detection framework designed for trustworthy and scalable knowledge discovery. Our system adaptively processes and aligns heterogeneous text and visual inputs, leveraging modality-specific Transformer models and a cross-modal attention mechanism to robustly capture nuanced bullying patterns. To ensure user privacy and accommodate distributed data silos, we integrate federated learning, enabling collaborative model improvement without exposing raw user data. Furthermore, our multitask learning scheme jointly optimizes cyberbullying detection and cross-modal relationship understanding, enhancing adaptability to evolving online behaviors. Extensive experiments on real-world social media datasets demonstrate that our approach consistently outperforms single-modal and simple fusion baselines, validating its effectiveness in adaptive, trustworthy, and privacy-preserving cyberbullying detection.