Large language models (LLMs) trained on extensive web and dialogue datasets risk inadvertently memorizing sensitive personal or copyrighted information. Regulatory frameworks, such as the EU’s General Data Protection Regulation (GDPR) “right to be forgotten” and the California Consumer Privacy Act (CCPA) mandate the immediate removal of such data, yet retraining the entire model incurs prohibitive computational and financial costs. Conventional gradient ascent-based unlearning methods apply uniform gradient updates across entire input sequences, significantly degrading overall language capabilities. In this study, we propose an unlearning framework that leverages a sparse autoencoder to selectively identify and remove only the most semantically critical tokens. On the TOFU $1 \%$ forget set, our method achieves a Forget Quality of 0.99 and a Model Utility of 0.61, improving performance by over $50 \%$ and $9 \%$, respectively, compared to standard approaches. This method minimizes unnecessary updates to preserve overall performance and enable more robust and efficient unlearning.

In this paper, we developed and experimented of the W-band 1-D focal plane array (FPA) passive imaging radiometer which was called by Millimeter-wave Imaging Radiometer Equipment (MIRAE). This consists of quasi-optics including a dielectric aspheric lens and a metal mirror as a reflector, 30 channels FPA receiver based on the direct conversion type, scanner, image processor and displayer. To get 2-D radiometer image with 1 Hz frame rate, the radiometric 1-D FPA receiver is moved into one direction by the linear scanner. From the experiment results, it is certain that MIRAE system can get enough good outdoor radiometer images to detect concealed metal weapons under clothing in the short range application such as security, comparing with other commercial and competitive passive imaging radiometer systems at W-band.

In this paper, we describe the adaptive hybrid watermarking scheme to improve the watermark detector's performance. In the hybrid watermarking, the conventional spatial watermarking and QIM scheme are applied simultaneously to the spatial domain of the image. According to the edge-map of the image the spatial watermarking is applied to the flat region and the QIM is applied to the high frequency region including edge and texture. The threshold between flat and edge is decided empirically and watermark blocks are embedded repeatedly into the image. In the detection process, we show how to estimate the threshold value by ACF(Auto-Correlation Function). Experimental results show the proposed hybrid watermarking estimates the watermark signal correctly than the conventional spatial watermarking.

Fingerprinting is one of many copyright protection methods and an extended technique of watermarking. In fingerprinting the identity of customer is embedded into the content in a way that is difficult to erase. The main difference between watermarking and fingerprinting is that different copies for each customer can be produced. Attackers compare several fingerprinting copies and find the location of the embedded information and destroy it by altering the values in those places where a difference was detected. In this paper, we propose a short fingerprinting code which is robust to the collusion attack of N customers, in the proposed fingerprinting scheme, two kinds of code are used. The first code has the same value for all customers and is embedded at the same location like watermark. The second code is designed unique for each customer. We apply this scheme to the digital image and show that it has some robustness against collusion attacks.

In recent years, digital watermarking has been introduced as a means of effectively protecting copyrights on digitized media such as image, audio and multimedia data. We present a watermarking technique for embedding digital seal images into a digital image, and apply this technique to each spatial and frequency domain of an image. Additionally, this technique does not require the original image to recover the seal image. This not only simplifies the encoding process but also makes up for the weak point which could be shown when an illegal attacker makes proving ownership difficult by creating a fake original. Experimental results show that the proposed watermarking technique is robust to various kinds of attacks.