The increasing heterogeneity and obfuscation of darknet markets present significant challenges for cross-market text classification. These markets often exhibit inconsistent syntax, domain-specific slang, and frequent use of encryption or code words. In light of this, we propose DarkFusionNet, a neural architecture for darknet text classification. It captures global semantics and local contextual features through Contextual Transformer Layer and Fusion Neural Network Layer. We also introduce DT-Dataset, a benchmark of annotated darknet texts collected from multiple markets. Experiments demonstrate the effectiveness of DarkFusionNet, surpassing existing baselines. Cross-market evaluations further confirm its robustness and generalizability to unseen markets with diverse linguistic styles and content.