Visual graph classification for blockchain security: Experiences fine-tuning Qwen2-VL on AMD MI300X [D]

Hi everyone, I’ve been working on a computer vision approach to a specific security problem in the "Agentic Economy": identifying malicious transaction patterns that are mathematically obfuscated but topologically distinct. The Problem Traditional rule-based security engines and even standard GNNs often struggle with "splitting attacks" - where a high-value transaction is fragmented into thousands of micro-transactions to bypass statistical thresholds.