Detecting of malicious logic in the operation of convolutional artificial neural networks based on the analysis of input data features
Jazyk: | ruština |
---|---|
Rok vydání: | 2023 |
Předmět: |
machine learning
vulnerability scanning convolutional neural networks malicious logic маÑинное обÑÑение поиÑк ÑÑзвимоÑÑей ÑвеÑÑоÑнÑе нейÑоннÑе ÑеÑи вейвлеÑ-пÑеобÑазование backdoor wavelet transform ÑкÑÑÑÐ°Ñ Ð»Ð¾Ð³Ð¸ÐºÐ° поÑайнÑе Ñ Ð¾Ð´Ñ |
DOI: | 10.18720/spbpu/3/2023/vr/vr23-828 |
Popis: | Тема вÑпÑÑкной квалиÑикаÑионной ÑабоÑÑ: «ÐбнаÑÑжение вÑедоноÑной логики в ÑабоÑе ÑвеÑÑоÑнÑÑ Â Ð¸ÑкÑÑÑÑвеннÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей на оÑнове анализа пÑизнаков Ð²Ñ Ð¾Ð´Ð½ÑÑ Ð´Ð°Ð½Ð½ÑÑ Â». ÐÑедмеÑом иÑÑÐ»ÐµÐ´Ð¾Ð²Ð°Ð½Ð¸Ñ ÑвлÑеÑÑÑ Ð·Ð°ÑиÑа клаÑÑиÑикаÑоÑов на оÑнове ÐÐС Ð¾Ñ Ð²Ð½ÐµÐ´ÑÐµÐ½Ð¸Ñ Ð¿Ð¾ÑайнÑÑ Ñ Ð¾Ð´Ð¾Ð², акÑивиÑÑÑÑÐ¸Ñ ÑкÑÑÑÑÑ Ð»Ð¾Ð³Ð¸ÐºÑ.ЦелÑÑ ÑабоÑÑ ÑвлÑеÑÑÑ Ð·Ð°ÑиÑа ÑвеÑÑоÑнÑÑ Â Ð¸ÑкÑÑÑÑвеннÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей Ð¾Ñ Ð°Ñак внедÑÐµÐ½Ð¸Ñ ÑкÑÑÑой логики. ÐадаÑи, ÑеÑаемÑе в Ñ Ð¾Ð´Ðµ иÑÑледованиÑ:ÐÑÑледование ÑеоÑеÑиÑеÑÐºÐ¸Ñ ÑпоÑобов ÑеализаÑии аÑак внедÑÐµÐ½Ð¸Ñ ÑкÑÑÑой логики.Ðнализ ÑовÑеменнÑÑ Ð¸ÑÑледований в облаÑÑи поиÑка ÑкÑÑÑой логики в ÑвеÑÑоÑнÑÑ Ð¸ÑкÑÑÑÑвеннÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑÑÑ .ÐÑÑледование ÑпоÑобов ÑжаÑÐ¸Ñ Ð¸Ð·Ð¾Ð±Ñажений Ð´Ð»Ñ ÑÐ´Ð°Ð»ÐµÐ½Ð¸Ñ Ð¿Ñизнаков.РазÑабоÑка меÑода поиÑка ÑкÑÑÑой логики в ÑвеÑÑоÑнÑÑ Â Ð¸ÑкÑÑÑÑвеннÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑÑÑ Ð½Ð° оÑнове ÑжаÑÐ¸Ñ Ð²ÐµÐ¹Ð²Ð»ÐµÑ-пÑеобÑазованием. ÐÑенка каÑеÑÑва ÑабоÑÑ Ñозданного меÑода поиÑка ÑкÑÑÑой логики. РÑезÑлÑÑаÑе ÑабоÑÑ Ð±Ñло ÑазÑабоÑано ÑÑедÑÑво поиÑка ÑкÑÑÑой логики в ÑабоÑе ÑвеÑÑоÑнÑÑ Â Ð¸ÑкÑÑÑÑвеннÑÑ Â Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей, а Ñакже пÑодемонÑÑÑиÑована ÑÑÑекÑивноÑÑÑ ÐµÐ³Ð¾ ÑабоÑÑ. ÐÑл Ñделан вÑвод о Ñом, ÑÑо длÑ поиÑка Ñ Ð²ÑедоноÑной логики в ÑабоÑе ÑвеÑÑоÑнÑÑ Ð¸ÑкÑÑÑÑвеннÑÑ Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей ко Ð²Ñ Ð¾Ð´Ð½Ñм даннÑм могÑÑ Ð¿ÑименÑÑÑÑÑ Ð°Ð»Ð³Ð¾ÑиÑÐ¼Ñ ÑжаÑÐ¸Ñ Ñ Ð¿Ð¾ÑеÑÑми. ÐолÑÑеннÑе ÑезÑлÑÑаÑÑ Ð¼Ð¾Ð³ÑÑ Ð±ÑÑÑ Ð¸ÑполÑÐ·Ð¾Ð²Ð°Ð½Ñ Ð² каÑеÑÑве оÑÐ½Ð¾Ð²Ñ Ð´Ð»Ñ Ð¿ÑоекÑиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ÑиÑÑем динамиÑеÑкого поиÑка ÑкÑÑÑой логики в ÑабоÑе ÑвеÑÑоÑнÑÑ Â Ð¸ÑкÑÑÑÑвеннÑÑ Â Ð½ÐµÐ¹ÑоннÑÑ ÑеÑей. The topic of the graduate qualification work is «Detecting of malicious logic in the operation of convolutional artificial neural networks based on the analysis of input data features». The subject of the study is the protection of classifiers based on artificial neural networks from the introduction of secret passages that activate malicious logic.The purpose of the study is protect convolutional artificial neural networks from malicious logic attacks implementation. The research set the following goals:Research of theoretical ways to implement malicious logic implementation attacks.Analysis of modern research in the field of search for malicious logic in convolutional neural networks.Investigation of ways to compress images to remove features.Development of a method for searching for malicious logic based on compression using a wavelet transform.Evaluation of the quality of the created method of searching for malicious logic.As a result of the work, a means of combating malicious logic in the operation of convolutional neural networks was developed, and the effectiveness of its work was demonstrated. It was concluded that lossy compression algorithms can be applied to input data to combat malicious logic in the operation of convolutional artificial neural networks. The results obtained can be used as a basis for designing systems for dynamic search of malicious logic in the operation of convolutional neural networks. |
Databáze: | OpenAIRE |
Externí odkaz: |