Abstract: This paper chronicles the multi-faceted strategies used to defeat the Break Our Watermarking System (BOWS) contest. By blending social engineering, heuristic manipulation, and game-theoretic optimization, the authors demonstrate that breaking a secure system is rarely just about the mathematics—it is about identifying the weakest link in the entire pipeline, from human behavior to "oracle" vulnerabilities.
Introduction
Digital watermarking is a technique used to embed hidden data into media for copyright protection. The BOWS contest challenged researchers to "break" a secret watermarking system applied to three unknown $512 \times 512$ grayscale images. The objective was to modify these images so that a secret detector could no longer identify the watermark, while maintaining a Peak Signal-to-Noise Ratio (PSNR) above 30 dB to ensure visual quality. The core difficulty lay in the "black box" nature of the system: the algorithm was secret, and contestants only had access to a binary online detector.
Methods: The Three Personas
The authors approached the problem using three distinct methodologies, categorized by the personas of a classic Spaghetti Western:
1. "The Bad" (Social Engineering)
Rather than attacking the bits directly, this strategy targeted the human element. By networking with developers at conferences and "socially hacking" information over drinks, the researchers confirmed the likely algorithm was based on a specific paper by Miller, Doërr, and Cox. Most significantly, they located the original, unwatermarked version of the "church" image in a previous publication by the steering committee. This allowed them to perfectly align and subtract the watermark from the contest image.
2. "The Ugly" (Heuristic Hacking)
This approach involved "Weird and Ugly Attacks" (WUA). Researchers manually manipulated Discrete Cosine Transform (DCT) coefficients—zeroing them out or randomly shuffling them—until the detector failed. This trial-and-error method prioritized breaking the detector over maintaining perfect image aesthetics, though it remained within the contest's quality constraints.
3. "The Good" (Theoretical Optimal Attack)
Assuming a spread-spectrum watermarking model, the authors formulated the problem as a game between an embedder and an attacker. Using a Lagrangian formulation, they identified the optimal noise addition (a Wiener filter-like approach) to minimize the detector's signal-to-noise ratio (SNR) while staying within the PSNR limits.
Dataset and Results
The study utilized the official BOWS contest dataset, consisting of three $512 \times 512$ grayscale images: a strawberry, a wood path, and a church.
The authors finished in 3rd place with a mean PSNR of 35.24 dB. The "Bad" attack on the church image yielded the highest quality result for that specific image (37.35 dB) because it leveraged the original source data. The contest was ultimately won by a team using an "oracle attack," which iteratively queried the detector to "carve" away the watermark with minimal distortion—a strategy the authors initially overlooked in favor of theoretical rigor.
Key Insights and Conclusion
The primary takeaway is that the human element and the "oracle" vulnerability of the detector were the most effective points of attack. While the theoretical "Good" attack provided a systematic framework, the "Bad" social engineering attack yielded superior results by exploiting flaws in the contest's setup, such as the reuse of images. This research highlights that "security through obscurity" is insufficient when faced with attackers who can exploit both human nature and mathematical vulnerabilities. Future security challenges must ensure that datasets are not reused and that detector queries do not inadvertently leak information.
Summary updated