Earticle

최근, 대형 언어 모델(Large Language Model)을 대상으로 혐오적, 폭력적, 또는 공격적인 표현을 프롬프트로 입력하여 모 델의 응답을 왜곡하거나 악용하는 사례가 발생하고 있다. 이러한 공격적 사용에 대비하기 위해, 본 논문에서는 이중 필터링 기법을 제안한다. 제안된 기법은 1단계에서 혐오적 문장 여부를 분류할 수 있는 분류기를 적용하고, 2단계에서 기존 지식 내 에서 벡터 검색을 활용하여 문장의 의미를 추가적으로 평가하여 잘못 분류된 문장을 걸러내는 과정을 포함한다. 이를 통해 모 델의 응답이 공격적 콘텐츠를 포함하지 않도록, 보다 정교한 필터링을 구현하여 언어 모델의 안전성을 강화할 수 있도록 한다. 이를 위해, 1단계에서는 텍스트 임베딩 모델과 텍스트 분류기를 결합하여 혐오적 문장 분류 실험을 수행하였다. 그 결과, 0.98 이상의 f-score를 보였다. 2단계에서는 텍스트 임베딩 모델을 활용하여 질문 및 답변을 벡터 값으로 전환하여 질문에 대한 답 을 찾는 실험을 수행하였다. 그 결과, 0.87의 Precision@1을 보였다.

Recently, there have been instances where prompts containing hateful, violent, or offensive expressions are input into large language models (LLMs) to distort or misuse their responses. To counteract such abusive usage, this paper proposes a dual filtering technique. The proposed method applies a classifier in the first stage to detect hateful sentences, followed by a second stage where vector search is employed to further assess the semantics of the sentence within the model's existing knowledge base, filtering out any misclassified sentences. This approach aims to strengthen the safety of language models by implementing more sophisticated filtering that prevents the inclusion of offensive content in model responses. In the first stage, we combined a text embedding model with a text classifier to conduct an experiment on classifying hateful sentences, achieving an f-score of over 0.98. In the second stage, we conducted experiments using a text embedding model to convert questions and answers into vector values, with a focus on finding accurate responses to queries. The results showed a Precision@1 of 0.87.

토렌트 및 P2P 사이트나 웹 하드는 쉽고 편리하게 무료로 다운로드 받거나 가격을 저렴하게 이용할 수 있다는 이유 만으로 사용자들은 자주 이용하지만 국내 토렌트 및 P2P 사이트나 웹 하드는 저작권과 관련되어 매우 민감하기 때문에 저작권 보호를 위한 기술들이 연구되고 적용 되어지고 있다. 이 중에서 파일의 제목이나 주요 단어의 조합 등 경우의 수를 금칙어로 설정하여 차단하는 제목 및 문자열 비교방식 필터링 기술은 제목 변경, 띄어쓰기 등을 통해 우회가 용이 하다. 저작권 보호를 위한 불법저작물을 검색하고 차단하기 위해서는 변형된 파일 제목을 정규화 하는 기술이 필수적이 다. 본 논문에서는 불법저작물의 변형된 파일 제목을 정규화 하는 기법과 파일 제목을 정규화를 진행 전과 후에 따른 검색에 의한 탐지율을 비교하였다. 정규화를 진행하기 전 탐지율은 77.72%로 아쉬운 탐지율이 보인 반면에 정규화를 진 행한 후 90.23%로 정규화가 필수적이라고 말할 수 있다. 향후, 공통으로 나타나는 날짜와 화질 표시 같은 무의미한 용어 들을 처리하면, 더욱 좋은 결과가 산출될 것으로 기대한다. 국문 요약입니다.

Although torrents and P2P sites or web hard are frequently used by users simply because they can be easily do wnloaded freely or at low prices, domestic torrent and P2P sites or web hard are very sensitive to copyright. Techni ques have been researched and applied. Among these, title and string comparison method filtering techniques that bl ock the number of cases such as file titles or combinations of key words are blocked by changing the title and spaci ng. Bypass is easy through. In order to detect and block illegal works for copyright protection, a technique for norm alizing modified file titles is essential. In this paper, we compared the detection rate by searching before and after no rmalizing the modified file title of illegal works and normalizing the file title. Before the normalization, the detection rate was 77.72%, which was unfortunate while the detection rate was 90.23% after the normalization. In the future, i t is expected that better handling of nonsense terms, such as common date and quality display, will yield better resu lts.

[목적] 광학식 캡처 장비를 활용한 골프 임팩트에 관한 3차원 영상 분석 시, 필터링은 자료 처리 과정의 일부이 며, 이때 클럽헤드의 속력을 제대로 산출하기 위해 적절한 차단 주파수 설정이 필요하다. [방법] 이를 확인하기 위해 대학 선수 4명에게 본인의 웨지 아이언으로 30m, 50m, 70m, 90m 거리를 목표로 샷을 하게 하여 각각의 다 른 클럽헤드 속력을 만들어 내었다. 이후, 획득한 좌표 자료로 파워 스펙트럼 분석을 하여 90%, 95%, 99%, 99.5% 의 누적 파워에 해당하는 차단 주파수를 산출하였고, 산출된 차단 주파수 값을 적용하여 클럽헤드의 속력을 재차 산출하였다. [결과] 연구 결과, 클럽헤드의 빠른 움직임 특성을 제대로 반영하기 위해서는 95~99% 누적 파워에 해 당하는 차단 주파수를 적용하여 필터링을 한 후, 속력을 산출하는 것을 권장하는데, 이는 낮은 차단 주파수 설정 시 원래 속력을 과소 추정하는 경향이 나타났기 때문이다. [결론] 추후 모션 캡처 시스템을 활용한 골프 영상 분 석 연구 시, 적절한 차단 주파수를 설정하여 원신호의 특성을 제대로 반영해야 할 것이며, 이때 파워 스펙트럼 분석을 활용하는 것을 권장한다.

[Purpose] In 3D motion analysis of golf performance using an optical capture device, filtering is fundamental part of the data processing, and it is necessary to carefully set an optimal cut-off frequency to calculate the club head’s speed accurately. [Method] To determine the appropriate cut-off frequency in golf shots being done at high speeds compared to general human movements, four university athletes shot for target distances of 30m, 50m, 70m and 90m with their wedge irons to generate different club head's speeds. The cut-off frequency corresponding to the cumulative power of 90%, 95%, 99% and 99.5% was calculated by the power spectral analysis with the raw coordinate data, and the speeds of the club head were calculated again by applying the relative cut-off frequency values. [Results] It might be recommended to calculate the speed after filtering by applying the cut-off frequency corresponding to 95 ~ 99% cumulative power in order to reflect the fast-moving characteristic of the club head properly. The tendency was to underestimate the veridical speed when applying a relatively low cut-off frequency. [Conclusion] It is necessary to properly apply the cut-off frequency in data filtering to properly reflect the characteristics of the original signal, and it would be recommended to use power spectral analysis.

대용량 모바일 기기의 발전과 보급이 확산됨에 따라 사용자들은 사진, 음악, 동영상과 같은 멀티미디어 콘텐츠를 대량으로 휴대하며 이용할 수 있게 되었다. 그러나, 이러한 대량의 멀티미디어 콘텐츠 관리는 사용자 각자에게 맡겨져 있어 콘텐츠 관리를 어렵게 하고 있는 현실이다. 본 논문에서는 분산 모바일 환경에서 멀티미디어 콘텐츠의 공유와 추전을 통해 사용자에게 적합한 콘텐츠를 추천을 통해 제공하고, 제공된 콘텐츠는 모바일 동기화 서비스를 통해 모바일 기기로 저장하고 관리되는 '분산 모바일 환경에서 멀티미디어 콘텐츠 추전 및 검색 서비스'를 설계하고 구현하였다. 제안된 시스템은 사용자의 선호 프로파일 정보로 협업 필터링을 통해 공유된 멀티미디어 콘텐츠 중에서 사용자에게 적합한 콘텐츠를 추천해 주고, 추천된 콘텐츠는 모바일 기기 사용자의 행동에 따라 모바일 동기화 서비스를 통해 모바일 기기에 저장과 관리, 검색이 된다. 본 논문에서 제안된 방법은 추천과 검색을 통해 사용자 모바일 기기의 멀티미디어 콘텐츠를 효율적으로 관리 할 수 있다. 이처럼 본 논문에서 제안된 서비스 방법은 멀티미디어 콘텐츠의 추천과 모바일 동기화 서비스로 능동적인 콘텐츠 관리를 제공하며, 사용자에게 효율적인 콘텐츠 검색 기법과 활용 방법을 제공 할 수 있다.

As the number of low Earth orbit satellites is increasing, new algorithms and high-performance computing resources are needed to handle the enormous real-time arrival data generated by satellites. Based on the requirements, this paper proposes a novel real-time adaptive speckle filtering selection algorithm for satellite synthetic aperture radar (SAR) images. The use of high-performance filters generates high-quality images whereas it introduces delays. Thus, a real-time adaptive algorithm which achieves time-average SAR image quality maximization subject to delays using Lyapunov optimization, where the delay is formulated with a queuing model. Evaluation results show that the proposed algorithm guarantees desired performance improvements.

In this paper, we present a convergence proof for an iterative procedure of local mode filtering. We formulate the local mode filtering as a quadratic optimization problem based on the Legendre transform of convex function, from which two closed-form expressions at each iteration step are derived for variables to be optimized. Those analytical solutions ensure that the value of objective function increases monotonically with the progress of the iterative procedure. We also show experimental results using a grayscale image, which support our theoretical results practically.

최근 주목을 받고 있는 Particle Filtering은 실제 객체 추적에서 발생하는 비선형, 비 가우시안 분포를 가지는 상태 벡터의 사후확률 을 추정하기 위한 Monte Carlo 시뮬레이션에 기반을 둔 추적 방법론이다. 우리는 본 논문에서 Particle Filtering을 이용한 객체 추적 성능을 향상시킬 수 있는 두 가지 방법론을 제안한다. 첫 번째는 확률이 가장 낮은 샘플을 이전 프레임의 추정된 상태 벡터로 대치하 는 피드백 방법론이고, 두 번째는 객체 확률 분포를 추정된 객체 후보영역에 역투영하여 신뢰구간을 구함으로써 추적 박스의 정확도 를 향상시키는 방법이다. 또한, 실험을 통해 구한 추적 샘플의 진화 방정식을 제시하였다. 우리는 다양한 상황이 설정된 실험 데이터 셋을 구성하여 실험을 실시하여 제안한 방법론의 우수성을 입증하였다.

The object tracking method using particle filtering has been proved successful since it is based on the Monte Carlo simulation to estimate the posterior distribution of the state vector that is nonlinear and non-Gaussian in the real-world situation. In this paper, we present two nobel methods that can improve the performance of the object tracking algorithm based on the particle filtering. First one is the feedback method that replace the low-weighted tracking sample by the estimated state vector in the previous frame. The second one is an tracking box correction method to find an confidence interval of back projection probability on the estimated candidate object area. An sample propagation equation is also presented, which is obtained by experiments. We designed well-organized test data set which reflects various challenging circumstances, and, by using it, experimental results proved that the proposed methods improves the traditional particle filter based object tracking method.

ChatGPT의 등장 이후 대규모 언어 모델(Large Language Model, LLM)을 다양한 분야에 적용하는 연구가 활발하게 이루어지고 있다. LLM의 추론 성능을 향상시키기 위한 대표적인 방법으로는 Chain-of-Thought, Retrieval-Augmented Generation(RAG) 등이 있다. 그러나 이러한 방법들은 입력 프롬프트(Prompt)의 길이를 증가 시켜 추론 시간과 비용을 높이며, 불필요한 문맥으로 인해 환각 현상이 발생하는 문제가 있다. 본 논문에서는 이러한 문제를 해결하기 위해, 프롬프트 중 질의문을 기준으로 관련된 정보만 남기고 문맥(Context)을 압축하는 문맥 필터링 (Context Filtering) 기법을 제안한다. 문맥 필터링은 질의문의 시계열적, 의미적 관련성을 기준으로 필요한 정보만 입력 으로 사용하여 LLM이 추론하는 방법이다. 실험은 전장 시뮬레이션으로부터 생성된 Timestamp + Triplet 구조의 대규 모 시계열 데이터를 기반으로 수행되었다. 실험 결과, 전체 문맥의 약 25%만을 사용한 압축 프롬프트에서도 기존 프롬프 트와 유사한 수준의 상황 인지 정확도를 유지하였으며, 추론 시간과 토큰 사용량 또한 감소하였음을 확인하였다.

Since the advent of ChatGPT, research on applying large language models(LLM) to various fields has been actively conducted. Typical methods for improving LLM's reasoning performance include Chain-of-Thought and Retrieval-Augmented Generation(RAG). However, these methods increase the length of the input prompt, thereby increasing the inference time and cost, and there is a problem that Hallucination occurs due to unnecessary context. In order to solve this problem, this paper proposes a Context Filtering technique that compresses context, leaving only relevant information based on the questionnaire among prompts. Context Filtering is a method for LLM to infer using only necessary information as input based on the time-series and semantic relevance of a query. The experiment was conducted based on time-series data of Timestamp + Triplet structure in a battlefield situational awareness scenario. As a result of the experiment, it was confirmed that the situation awareness accuracy similar to that of the existing prompt was maintained even at the compression prompt using only about 25% of the entire context, and the inference time and token usage were also reduced.

전통적인 회선교환 방식에서 패킷 교환 방식으로 네트워크가 진화함에 따라 분산된 클럭의 동기화를 위한 패킷 기반 동기화 방식은 큰 주목을 받고 있다. 특히 네트워크의 혼잡이 심각한 경우 패킷교환 네트워크의 특성에 기인하는 패킷 지연 변화 (PDV)는 IEEE1588 PTP (Precision Time Protocol)를 기반으로 하는 패킷 기반의 동기화 시스템에서 클럭 잡음의 주요 원인으로 작용한다. 본 논문에서는 IEEE1588 PTP의 동기화 성능을 개선하기 위해 이러한 클럭 잡음을 완화시킬 수 있는 비선형 SMoPF (Sample Mode Packet Filtering) 메커니즘을 제안하였다. 제안된 SMoPF 기법의 성능평가를 얻기 위하여 임의의 혼잡도를 갖는 패킷 교환 네트워크에서의 IEEE1588 PTP 동기화 프로토콜을 고려하였으며OMNET++ 툴 기반의 컴퓨터 시뮬레이션을 수행하였다. 또한, 제안된 기법의 성능은 기존의 샘플 최소 필터링, 평균 필터링, 최대 필터링 기법들과 비교 분석 되었다.

Packet-based methods for synchronizing distributed clocks are getting tremendous attention as networks are evolving from the conventional circuit-switched to packet-switched architecture. Packet Delay Variation (PDV) inherent in packet-switched networks, especially when the network is moderately or heavily congested, is a dominant source of clock noise in the packet-based synchronization systems that are based in IEEE 1588 Precision Time Protocol (PTP). This paper presents a non-linear Sample Mode Packet Filtering (SMoPF) mechanism to smooth such a clock noise to enhance the synchronization performance of IEEE 1588 PTP. Performance of the proposed SMoPF mechanism is compared (via computer simulations in OMNET++) with that of existing sample minimum-, sample mean-, and sample-maximum filtering mechanisms considering the IEEE 1588 PTP synchronization domain on top of an arbitrarily congested packet-switched network.

The in-vehicle infotainment (IVI) system provides driving-related information such as vehicle status and route guidance, as well as entertainment services for users. In other words, the IVI system performs a variety of functions such as audio and video data streaming, navigation, smartphone pairing and mirroring, and voice-activated vehicle control. Therefore, the IVI system stores a very large amount of log data, including fuel consumption, current vehicle status, navigation information, as well as information about external devices (e.g., smartphones, tablets) connected to it. Manually analyzing such log data is cumbersome and time-consuming. To solve this problem, this paper proposes a method to efficiently analyze log data of the IVI system. The proposed method can automatically extract and analyze important information, enabling effective vehicle digital forensic investigation.

As the rapid growth of mobile device, a recommender system is required to provide adequate recommendation list even if the customers provide their needs explicitly. In this research, the problem that customers provide their needs is modeled as constraint satisfaction problem. However the existing constraint satisfaction is too rigid, so we employ approximate constraint satisfaction by adopting indifference interval. The proposed recommendation methodology is composed of two phases; the first phase is related to CF-based filtering to generate the candidate recommendation set. The second phase is related to approximate constraint filtering to find the final adequate items fitting individual customers’ concerns. We expect that the proposed methodology contribute to display the adequate items to customer’s concerns for better recommendation in mobile environment.

본 논문은 Extrapolative Collaborative Filtering 방법론의 유용성을 컨텐트 추천 분야에서 확인하고자 한다. 컨텐트 서비스들은 개별적으로 운영되고, 그들간의 정보 공유가 이루어지지 않아, 타 서비스를 이용할 경우 다른 서비스에서의 자신의 성향이 반영되지 않는 한계가 있다. 이에 본 연구에서는 다양한 컨텐트 사이트를 이용하는 사용자의 편의를 위해 타 사이트간의 최소한의 협력이 있는 환경에서, 상호적 추천을 하는 Extrapolative Collaborative Filtering (ECF) 방법론이 유용한지를 검증하였다. 이를 위해 공개된 4개의 컨텐트 사이트 데이터셋을 확보하여 실험하였다. 그 결과, 적절한 협력을 하는 경우가 독자적으로 하는 것보다 추천의 성과가 향상될 수 있음을 확인하였다.

협력 필터링은 상업적 추천 시스템에서 널리 사용되어 왔는데, 고객의 사회적 행태를 구현하여 사용자의 흥 미에 부합하는 항목들을 제안하기 때문이다. 현재까지 적절한 항목을 추천하기 위한 가장 보편적인 방법은 유사한 사용자들을 찾아 그들의 평가치를 참조하는 방법이다. 본 논문은 영화를 추천하기 위해서 장르 흥미도를 기반으로 하는 새로운 유사도 공식을 제안하는데, 이는 기존 공식에서 사용자들의 평가등급 차이를 기반으로 하는 것과 대비 된다. 광범위한 실험 결과에 따르면, 제안한 공식은 정확도와 추천의 질에 있어서 전통적인 유사도 공식의 성능을 크게 향상시키는 것으로 확인되었다.

Collaborative filtering has been popular in commercial recommender systems, as it successfully implements social behavior of customers by suggesting items that might fit to the interests of a user. So far, most common method to find proper items for recommendation is by searching for similar users and consulting their ratings. This paper suggests a new similarity measure for movie recommendation that is based on genre interest, instead of differences between ratings made by two users as in previous similarity measures. From extensive experiments, the proposed measure is proved to perform significantly better than classic similarity measures in terms of both prediction and recommendation qualities.

The proliferation of items increased the difficulty of customers in finding the specific items they want to purchase. To solve this problem, companies adopted recommender systems, such as collaborative filtering systems, to provide personalization services. However, companies use only meaningful and essential data given the explosive growth of data. Some customers are concerned that their private information may be exposed because CF systems necessarily deal with personal information. Based on these concerns, we analyze the effects of the amount of information on recommendation performance. We assume that a customer could choose to provide overall information or partial information. Experimental results indicate that customers who provided overall information generally demonstrated high performance, but differences exist according to the characteristics of products. Our study can provide companies with insights concerning the efficient utilization of data.

In this paper, we propose a method to automatically generate a height map image that can be useful for creating 2.5D clip. To this end, we used a Viola - Jones algorithm that segments eyes, nose and mouth of the image and analyzed the characteristics of each to derive the height value that is used for producing a final height map . Finally, we can apply the height map to create photo-realistic images that can be easily reproduced for synthesizing 2.5D animation in Adobe. To verify the proposed method, we have carried out a user study between 2.5D movies clips which are proposed our automatic method and a manually generated method.

With the development of related technologies, Location-Based Services (LBS) are growing fast and being used in many ways. Past LBS studies have focused on adoption of LBS because of the fact that LBS users have privacy concerns regarding revealing their location information. Meanwhile, the number of LBS users and revenues from LBS are growing rapidly because users can get some benefits by revealing their location information. Little research has been done on how LBS affects consumers’ information search behavior in product purchase. The purpose of this paper is examining the effect of LBS information filtering on buyers’ uncertainty and their information search behavior. When consumers purchase a product, they try to reduce uncertainty by searching information. Generally, there are two types of uncertainties – knowledge uncertainty and choice uncertainty. Knowledge uncertainty refers to the lack of information on what kinds of alternatives are available in the market and/or their important attributes. Therefore, consumers having knowledge uncertainty will have difficulties in identifying what alternatives exist in the market to fulfil their needs. Choice uncertainty refers to the lack of information about consumers’ own preferences and which alternative will fit in their needs. Therefore, consumers with choice uncertainty have difficulties selecting best product among available alternatives.. According to economics of information theory, consumers narrow the scope of information search when knowledge uncertainty is high. It is because consumers’ information search cost is high when their knowledge uncertainty is high. If people do not know available alternatives and their attributes, it takes time and cognitive efforts for them to acquire information about available alternatives. Therefore, they will reduce search breadth. For people with high knowledge uncertainty, the information about products and their attributes is new and of high value for them. Therefore, they will conduct searches more in-depth because they have incentive to acquire more information. When people have high choice uncertainty, people tend to search information about more alternatives. It is because increased search breadth will improve their chances to find better alternative for them. On the other hand, since human’s cognitive capacity is limited, the increased search breadth (more alternatives) will reduce the depth of information search for each alternative. Consumers with high choice uncertainty will spend less time and effort for each alternative because considering more alternatives will increase their utility. LBS provides users with the capability to screen alternatives based on the distance from them, which reduces information search costs. Therefore, it is expected that LBS will help users consider more alternatives even when they have high knowledge uncertainty. LBS provides distance information, which helps users choose alternatives appropriate for them. Therefore, users will perceive lower choice uncertainty when they use LBS. In order to test the hypotheses, we selected 80 students and assigned them to one of the two experiment groups. One group was asked to use LBS to search surrounding restaurants and the other group was asked to not use LBS to search nearby restaurants. The experimental tasks and measures items were validated in a pilot experiment. The final measurement items are shown in Appendix A. Each subject was asked to read one of the two scenarios – with or without LBS – and use a smartphone application to pick a restaurant. All behaviors on smartphone were recorded using a recording application. Search breadth was measured by the number of restaurants clicked by each subject. Search depths was measured by two metrics – the average number of sub-level pages each subject visited and the average time spent on each restaurant. The hypotheses were tested using SPSS and PLS. The results show that knowledge uncertainty reduces search breadth (H1a). However, there was no significant correlation between knowledge uncertainty and search depth (H1b). Choice uncertainty significantly reduces search depth (H2b), but no significant relationship was found between choice uncertainty and search breadth (H2a). LBS information filtering significantly reduces the buyers’ choice uncertainty (H4) and reduces the negative relationship between knowledge uncertainty and search breadth (H3). This research provides some important implications for service providers. Service providers should use different strategies based on their service properties. For those service providers who are not well-known to consumers (high knowledge uncertainty) should encourage their customers to use LBS. This is because LBS would increase buyers’ consideration sets when the knowledge uncertainty is high. Therefore, less known services have chances to be included in consumers’ consideration sets with LBS. On the other hand, LBS information filtering decrease choice uncertainty and the near service providers are more likely to be selected than without LBS. Hence, service providers should analyze geographically approximate competitors’ strength and try to reduce the gap so that they can have chances to be included in the consideration set.