Imagine effortlessly copying the "brains" of a powerful AI model just by interacting with it. Sounds like sci-fi, but it's a real threat to companies offering Embeddings-as-a-Service (EaaS). These services, which provide text embeddings (numerical representations of text) generated by Large Language Models (LLMs) like OpenAI's GPT-3, are vulnerable to imitation attacks. Essentially, malicious actors can query the service, collect the output embeddings, and train their own "knock-off" model, bypassing service fees and potentially even offering competing services. To combat this, researchers have developed watermarking techniques to protect the intellectual property of these valuable models. Think of it like a hidden signature within the embeddings themselves. One method, called EmbMarker, inserts a specific target embedding into the original embedding, triggered by certain words in the input text. However, this technique has weaknesses. A clever attacker can identify and remove these watermarks, rendering them useless. An improved method, WARDEN, uses multiple target embeddings, making detection more robust. But even WARDEN has a vulnerability: paraphrasing. By slightly rewording the input text and averaging the resulting embeddings, attackers can dilute the watermark signals, making them hard to detect. This highlights a critical challenge: protecting embeddings while preserving their usefulness for legitimate users. The core problem lies in the trigger word mechanism. Because current techniques rely on specific words to activate the watermark, they are susceptible to manipulation through paraphrasing, which disrupts the trigger words and weakens the watermark signal. A new technique called WET (Watermarking EaaS with Linear Transformation) offers a solution. Instead of relying on trigger words, WET uses a secret linear transformation matrix to modify the entire embedding. To verify if a model has been copied, the inverse transformation is applied. If the watermark is present, the recovered embedding should closely resemble the original. This approach is robust to paraphrasing because the watermark is embedded in the fundamental structure of the embedding, not tied to specific words. Extensive tests show that WET provides near-perfect verifiability while preserving the utility of the embeddings for downstream tasks. This breakthrough marks a significant step in ensuring the security and integrity of EaaS offerings. The development of WET shows that watermarking can be an effective strategy for protecting EaaS. However, challenges remain, particularly around designing more sophisticated transformation matrices to make watermarks even harder to reverse engineer. As AI becomes more pervasive, protecting these valuable models will only become more critical. Techniques like WET represent the cutting edge of this ongoing effort, ensuring that the benefits of AI are accessible while safeguarding the intellectual property of its creators.