This paper presents a novel deep learning-based method for detecting robust statistical arbitrage strategies in financial markets, addressing model ambiguity. The authors define robust statistical arbitrage as strategies that yield positive expected profits conditional on terminal asset prices under all probability measures in an ambiguity set. They construct the ambiguity set using a Wasserstein ball around the empirical distribution of historical returns, making the approach entirely data-driven and model-free. A key theoretical contribution is Theorem 2.14, which shows that minimizing a penalized functional over neural network-represented trading strategies is equivalent to finding robust statistical arbitrage strategies for a sufficiently large penalty parameter. This allows the method to scale to high-dimensional settings (e.g., 50 assets) where traditional linear programming approaches fail. Empirically, the authors test their method on S&P 500 data and demonstrate that the identified strategies generate significant profits even during financial crises and when cointegration relationships break down, outperforming classical pairs trading. The results highlight the method's robustness and practical applicability in real-world markets.