ZAPs: A Reward Attribution Framework for DeFi Ecosystems with Adversarial-Robust Scoring via Parallel Anomaly Ensemble Detection

Battle Count: 49

The paper is primarily about DeFi incentive design and reward attribution rather than trading strategy or market microstructure. However, it has indirect relevance: (1) understanding volume semantics and the divergence between nominal volume and economic contribution informs trading signal quality; (2) sybil detection and adversarial robustness techniques are transferable to detecting wash trading and artificial volume in markets; (3) the percentile-normalization and cross-domain weighting concepts could inform factor construction; (4) the anomaly detection ensemble methodology is applicable to detecting market manipulation patterns. The paper does not address price prediction, portfolio optimization, or execution strategies directly.

The system involves multiple interacting components: percentile-normalized scoring across protocols, nested cross-domain weighting requiring sector classification, a four-layer detection stack (transaction gates, autoencoder + isolation forest ensemble, post-distribution memory, graph-based clustering), graduated penalty multipliers, and a behavioral quality model (zScore). The autoencoder architecture is simple (10→64→16→64→10), but the full pipeline requires: multi-chain transaction ingestion, protocol-to-sector classification, per-(wallet,protocol) decomposition, graph construction for funding provenance, coefficient-of-variation computation across families, and careful training-regime management for the isolation forest. The computational cost is manageable (1.77ms per wallet mean), but the system design complexity, parameter calibration, and adversarial robustness requirements make production deployment non-trivial.

The paper provides detailed architecture specifications (autoencoder dimensions, training hyperparameters like Adam lr=1e-3, batch=128, 50 epochs, 15% validation split), mathematical formulations, and algorithm pseudocode. However, critical parameters are explicitly withheld: mixing coefficients (α, β, γ), ensemble fusion weight λ (only λ*=0.90 reported), numeric penalty breakpoints, quality model calibration details, and the zScore behavioral engine internals. The labeled malicious corpus (1,073 wallets) and production data are not stated as publicly available. The protocol-to-sector classification mapping is proprietary. Full reproduction of production results is not possible without access to Zeru Finance's infrastructure and data.