Ensemble Boosting-Based Source Code Obfuscation Using Deep Learning Models for .NET

Abstract

Protecting. NET software against reverse engineering remains difficult because managed assemblies preserve rich structural information in Intermediate Language (IL) and metadata. This paper proposes an ensemble boosting-based obfuscation approach for. NET assemblies that combines three complementary deep learning components: an LSTM module for sequential IL dependencies, an autoencoder for latent structural transformation, and a transformer for global context modeling. Unlike static heuristic obfuscators, the proposed method selects transformation strategies adaptively at fragment level and aggregates model outputs through a boosting meta-stage. To ensure experimental reproducibility without relying on proprietary industrial software, a synthetic benchmark of managed-code fragments is introduced. The benchmark is generated from C# templates and evaluated at IL-token level using six obfuscation strategy classes: renaming, control-flow transformation, opaque predicate insertion, string encryption, dead-code insertion, and metadata-oriented protection. Experiments show that the boosting ensemble outperforms individual base models in transformation-selection accuracy. Compared with static baselines, the proposed method yields a more balanced trade-off among complexity growth, entropy increase, structural diversity, and overhead.