The History and Evolution of Watermarks — From Paper to AI
The history of watermarks dates back to 13th-century Italy. What began as patterns impressed on paper has evolved dramatically into watermarking in the modern era. This article traces the complete history from physical watermark origins to the latest deep learning-powered technologies.
The Origins — Paper Watermarks
The history of watermarks began in 1282 in Fabriano, Italy, where a paper mill impressed patterns onto paper. By using metal molds during the papermaking process, subtle differences in paper thickness created patterns visible when held up to light. This technique was used as trademarks for paper manufacturers and as marks of quality assurance. Eventually, watermarks were applied to prevent counterfeiting of banknotes, first adopted in 17th-century Swedish bank notes. Modern banknotes still employ this centuries-old watermarking technology. The portrait and pattern watermarks found in Japanese banknotes are extensions of this traditional technique. The key characteristic of physical watermarks is that they are integral to the paper itself, making removal or duplication extremely difficult. This 'difficult to remove' property became an important design goal for later watermarking technologies.
The Birth of Watermarks
Research into watermarking began in earnest in the early 1990s. In 1993, Tirkel et al. published a method for embedding information in images, and in 1994, Tanaka, Nakamura, and Matsui published papers using the term 'watermarking.' The rise of watermarking was driven by the rapid spread of the internet. Since digital content can be perfectly copied, new means of copyright protection were needed. Traditional encryption and DRM are effective for access control but cannot address unauthorized copying after content is decrypted. Watermarking was groundbreaking in that it embeds rights information directly into the content, maintaining it even after decryption. By the late 1990s, commercial products began to appear, with Digimarc (founded in 1995) providing the first commercial watermarking technology.
Three Eras of Watermarking
First Generation: Spatial Domain Methods (Early 1990s)
The earliest watermarking techniques used simple methods that directly manipulated pixel values. LSB (Least Significant Bit) substitution was the primary example, replacing each pixel's least significant bit with watermark information. While easy to implement and fast to process, it was vulnerable to even basic image processing like JPEG compression and noise addition. This era's technologies were mainly used for research and were insufficient for practical copyright protection.
Second Generation: Frequency Domain Methods (Late 1990s–2010s)
Frequency domain methods, appearing in the late 1990s, used DCT (Discrete Cosine Transform) or DWT (Discrete Wavelet Transform) to convert images to frequency components and embed information in their coefficients. Since JPEG compression is DCT-based, DCT-domain embedding offered high resistance to JPEG compression. DWT methods enabled multi-resolution analysis for more flexible embedding. This generation achieved practical-level robustness and saw adoption in commercial products. Digimarc and Philips commercialized technologies from this era.
Third Generation: Deep Learning Methods (2017–Present)
Around 2017, deep learning-powered watermarking technologies emerged. Neural networks automatically learn image features and determine optimal embedding positions and strengths. HiDDeN (2017), StegaStamp (2019), and Adobe TrustMark (2023) have successively pushed performance boundaries. The key advantage of deep learning methods is the ability to automatically acquire robustness against various attacks through training data. They achieve a combination of high quality preservation and robustness that was difficult with hand-designed algorithms.
Key Watermarking Technologies Through History
LSB Substitution (Early 1990s)
The most basic method that replaces the least significant bit of pixel values with watermark information. Extremely easy to implement but destroyed by even minor image processing. Today it is mostly used for steganography (secret communication) rather than copyright protection.
DCT/DWT Methods (Late 1990s–)
Methods that transform images to the frequency domain and embed watermark information in mid-frequency coefficients. Highly compatible with JPEG compression, and many commercial products are based on this technology. However, resistance to geometric transformations (rotation, scaling) was limited.
Deep Learning Methods (2017–)
Using encoder-decoder neural networks that learn image features for optimal embedding. HiDDeN (2017) was the pioneer, followed by RedMark (2018), StegaStamp (2019), MBRS (2022), and Adobe TrustMark (2023), which achieved significantly superior performance. TrustMark excels in both quality preservation (PSNR > 40dB) and robustness (bit accuracy > 98%).
The Present and Future of Watermarks
Currently, watermarking is taking on a new role in identifying AI-generated content. Google's SynthID, Meta's Stable Signature, and OpenAI's DALL-E watermark are examples of AI companies incorporating watermarking into their generative AI. The C2PA (Coalition for Content Provenance and Authenticity) standard combines watermarking with metadata for content authentication, with camera manufacturers (Sony, Nikon) and software companies (Adobe, Microsoft) announcing support. Future developments include expansion to video and audio, real-time processing acceleration, multimodal watermarking across media types, and research into quantum-resistant watermarking technologies.
Summary
Watermarks evolved from 13th-century paper watermarks through LSB methods, frequency domain methods, and deep learning methods. With the arrival of the AI era, proving content authenticity and provenance has become more important than ever. Adobe TrustMark, used by truvis, represents the cutting edge of this long history.