As online communication becomes increasingly essential in everyday life, the risk of falling victim to digital scams has grown significantly. These scams are no longer limited to single deceptive messages; In reality, the number of digital scams is growing by the day. These scams are no longer solo deceptive one-time messages but rather involve a full conversation in which the manipulative spirit slowly exerts its influence on the users. In such cases, either the detection systems are not able to pick up contextual information or they are unable to follow ever-evolving patterns of language. The thesis addresses an AI-within approach beyond a single message exchange on which one can base their scam detection: synthetic data augmentation for better performance. 

Background: Online-scams-enabled types have now become more dynamic and cases harder to detect, especially when unfolding over multiple interactions. Rule-based and classical machine-learning-based systems do not cope well with the mutable nature and cues of conversational scams. 

Objectives: The objective of this thesis is to augment the aim of scamming detection systems with Transformer-based models for grasping the context of messages. By focusing on RoBERTa and examining whether synthetically generated scam conversations can serve as a data augmentation mechanism against rare real data to enhance model performance. 

Methods: A three-stage procedure was run: first, synthetic scam conversations were generated through language models to augment the training data. The Classifiers Machine learning and deep learning, including RoBERTa, were trained on this data; and third, model performance was measured on a mixture of real and synthetic samples. 

Results: Trained with synthetic augmentation, RoBERTa beat the classically used classifiers in recognizing the scam patterns that emerge in conversations. The results suggest that combining synthetic and real data helps the model to detect the subtle and evolving tactics of scam. 

Conclusions: This thesis demonstrates the use of context-aware models coupled with synthetic data as a scalable and effective approach towards identifying multiturn scams. This approach is a step toward creating safer digital communication systems and highlights the prospects synthetic augmentation has in low-data cybersecurity contexts.