Search Results

Bots in Web3 Play-to-Earn (P2E) games extract on-chain rewards at scale, accelerating token inflation and eroding the incentives that sustain legitimate participation. Existing defenses transfer poorly to this setting. Traditional game-bot detectors require proprietary server-side telemetry that is unavailable on permissionless chains, and blockchain-native detectors target DeFi behaviors whose footprints differ fundamentally from those of game automation; research specific to Web3 game bots remains scarce, and no public labeled dataset exists. We argue that the very economic asymmetry that makes P2E automation profitable also exposes it, because the temporal, topological, financial, and behavioral regularities that bots cannot disguise without forfeiting their efficiency advantage remain observable from public on-chain transactions alone. Building on this insight, we propose a cross-game, cross-chain detection framework that models player activity as a heterogeneous temporal transaction graph over a unified five-type edge vocabulary, extracts 32 chain-agnostic behavioral features, and classifies accounts through a dual-branch encoder fusing a typed-edge graph attention network (GAT) with a gated recurrent unit (GRU)-attention sequence model, bootstrapped by graph-contrastive pre-training. To support evaluation, we further release the first multi-game labeled bot dataset, spanning three P2E titles on three heterogeneous chains with 57,879 accounts and 3,551 confirmed bots. On the joint test split, the framework attains F1 = 94.39%, AUC = 99.92%, and AP = 98.90%, with a sharply bimodal score distribution and stable performance across a wide range of decision thresholds.