• 정보보호학회논문지
• /
• 제34권5호
• /
• pp.981-991
• /
• 2024

인공지능 기술의 오작동을 유도하는 적대적 공격은 다양한 도메인과 모델에 적용 가능하며, 성능이 높은 SOTA(State-of-the-Art) 모델의 성능도 손쉽게 저해 시킬 수 있다. 이에 대처하기 위해 적대적 방어 기술들이 개발되고 있지만, 명확한 한계점으로 인해 활용이 제한된다. 그 결과, 특정 분야에서 인공지능 기술의 도입 뿐만 아니라 고도화 연구도 지체되고 있다. 해당 문제를 해결하기 위해, 본 논문에서는 적대적 공격의 손실 함수 업데이트 방향의 부호를 바꿔 새로운 개념의 적대적 데이터를 소개한다. 본 연구에서 소개한 역방향 적대적 데이터를 데이터 오염 및 적대적 훈련 환경에 적용하여 실험을 진행한 결과, 모델의 성능을 최대 72% 낮추고 9개 환경 중 6개 환경에서 강건성 향상에 가장 효율적임을 입증했다. 결과적으로, 제시한 적대적 데이터는 적대적 공격과 방어 기술의 연구 확장을 유도할 수 있으며, 더 나아가 방어 기술 개발의 고도화를 촉진할 수 있어 AI의 안전한 도입에 기여한다.

• 한국CDE학회논문집
• /
• 제11권5호
• /
• pp.335-343
• /
• 2006

Design changes for an original surface model are frequently required in a manufacturing area: for example, when the physical parts are modified or when the parts are partially manufactured from analogous shapes. In this case, an efficient 3D model updating method by locally adding scan data for the modified area is highly desirable. For this purpose, this paper presents a new procedure to update an initial model that is composed of combinatorial triangular facets based on a set of locally added point data. The initial surface model is first created from the initial point set by Tight Cocone, which is a water-tight surface reconstructor; and then the point cloud data for the updates is locally added onto the initial model maintaining the same coordinate system. In order to update the initial model, the special region on the initial surface that needs to be updated is recognized through the detection of the overlapping area between the initial model and the boundary of the newly added point cloud. After that, the initial surface model is eventually updated to the final output by replacing the recognized region with the newly added point cloud. The proposed method has been implemented and tested with several examples. This algorithm will be practically useful to modify the surface model with physical part changes and free-form surface design.

• 전자공학회논문지S
• /
• 제34S권2호
• /
• pp.52-62
• /
• 1997

To reconstruct the complete 3-D shape of an object, seveal range images form different viewpoints should be merged into a single model. The process of extraction of the transformation parameters between multiple range views is calle dregistration. In this paper, we propose a new algorithm to find the transformation parameters between multiple range views. Th eproposed algorithm consists of two step: initial estimation and iteratively update the transformation. To guess the initial transformation, we modify the principal axes by considering the projection effect, due to the difference fo viewpoints. Then, the following process is iterated: in order to extract the exact transformation parameters between the range views: For every point of the common region, find the nearest point among the neighborhood of the current corresponding point whose correspondency is defined by the reverse calibration of the range finder. Then, update the transformation to satisfy the new correspondencies. In order to evaluate the performance the proposed registration algorithm, some experiments are performed on real range data, acquired by space encoding range finder. The experimental results show that the proposed initial estimation accelerate the following iterative registration step.

• 한국정보디스플레이학회:학술대회논문집
• /
• 한국정보디스플레이학회 2004년도 Asia Display / IMID 04
• /
• pp.343-347
• /
• 2004

As a first demonstration of proven stability for a-Si, over 7000h of stability data for a high drive current a-Si AMOLED pixel driver circuit was presented at SID 2004. Also demonstrated was an AMOLED display operating at 6V using the IGNIS 4-T pixel driver circuit [1] and based on a top-emission reverse OLED architecture. In this paper, an update to that information is presented with test data that is representative of lifetime in excess of 25,000 hours taking into account pertinent acceleration factors, extracted from high current drive and high temperature stressing conditions. Lifetimes will continue to increase, given the high and ever-increasing OLED efficiency.

• International Journal of CAD/CAM
• /
• 제12권1호
• /
• pp.9-19
• /
• 2012

Generation of optimum planar B-spline curve in terms of minimum deviation and required fairness to approximate a target shape defined by a strip-shaped unorganized 2D point cloud is studied. It is proposed to use the location of control points as variables within the geometric optimization framework of point distance minimization. An adaptive simulated annealing heuristic optimization algorithm is developed to iteratively update an initial approximate curve towards the target shape. The new implementation comprises an adaptive cooling procedure in which the temperature change is adaptively dependent on the objective function evolution. It is shown that the proposed method results in an improved convergence speed when compared to the standard simulated annealing method. A couple of examples are included to show the applicability of the proposed method in the surface model reconstruction directly from point cloud data.