現実世界での機械学習によるアノマリ検知システムの構築と回避 by Clarence Chio - CODE BLUE 2015
Report
•6 likes•9,878 views
![199.72.81.55 - - [01/Jul/1995:00:00:01 -0400] "GET /history/apollo/ HTTP/1.0" 200 6245
unicomp6.unicomp.net - - [01/Jul/1995:00:00:06 -0400] "GET /shuttle/countdown/ HTTP/1.0" 200 3985
199.120.110.21 - - [01/Jul/1995:00:00:09 -0400] "GET /shuttle/missions/sts-73/mission-sts-
73.html HTTP/1.0" 200 4085
burger.letters.com - - [01/Jul/1995:00:00:11 -0400] "GET /shuttle/countdown/liftoff.html HTTP/
1.0" 304 0
199.120.110.21 - - [01/Jul/1995:00:00:11 -0400] "GET /shuttle/missions/sts-73/sts-73-patch-
small.gif HTTP/1.0" 200 4179
burger.letters.com - - [01/Jul/1995:00:00:12 -0400] "GET /images/NASA-logosmall.gif HTTP/1.0" 304 0
burger.letters.com - - [01/Jul/1995:00:00:12 -0400] "GET /shuttle/countdown/video/
livevideo.gif HTTP/1.0" 200 0
205.212.115.106 - - [01/Jul/1995:00:00:12 -0400] "GET /shuttle/countdown/countdown.html HTTP/
1.0" 200 3985
この中からどのようにアノマリを見つける？？？](https://image.slidesharecdn.com/makingbreakingmachinelearninganomalydetectorsinreallifej-160318081033/75/by-clarence-chio-code-blue-2015-8-2048.jpg?cb=1667651483)
現実世界での機械学習によるアノマリ検知システムの構築と回避 by Clarence Chio - CODE BLUE 2015
•6 likes•9,878 views
Download to read offline
Report
過去10年間で、機械学習(Machine Learning: ML)ベースのネットワーク侵入検知技術はWebセキュリティ業界において注目に値する勢いを獲得しました。一部の侵入検知システム(Intrusion Detection System: IDS)はこの技術を利用して侵入を検出し、ネットワークサービスに影響を与えさせることなく弾くことに成功しています。MLベースのIDSについて簡単に説明すると、IDSはまずWebのアクセスログを入力として正常なトラフィックのモデルを構築します。そして、オンライン処理システムにより正しいトラフィックの情報と不正なトラフィックの情報をモデルに学習させていきます。IDSはモデルと比較してある閾値以上離れたトラフィックを不正なものとして判定します。この理論では、システムに与えられるデータが多いほどモデルが正確になるため、進化・成長を続けるWebのトラフィックパターンを解析する上では一見して万能なシステムであるかのように思われます。 しかし、このお伽話はそう長くは続きませんでした。この分類モデルに対して、攻撃者が「ポイズニング」を行うことで侵入検知を回避していることが分かったのです。攻撃者はそれ自体は無害なトラフィックを大量に送り、攻撃ベクトルが閾値内に収まるように少しずつモデルを学習させるという方法で攻撃に成功しました。 この講演では、この「モデルポイズニング」攻撃のデモを行うと共に、MLベースの異常検知システムを攻撃者の操作による影響を受けにくくするための手法を解説します。背景にあるMLの理論そのものよりも、実世界で動いているこういったIDSやそれらを無効化する攻撃、そしてネットワーク侵入検知に関わる開発者への影響についてお話しします。さらに、最も重要な点としてMLベースのネットワーク侵入検知の未来についてお話しします。
Recommended
「のどが渇いた」というユーザーに何を出す？ ユーザーの「欲しい」に惑わされない、本当のインサイトを見つけるUXデザイン・UXリサーチYoshiki Hayama
54.7K views•243 slides
More Related Content
What's hot
What's hot(20)
More from CODE BLUE
More from CODE BLUE(20)
Recently uploaded
Recently uploaded(10)
現実世界での機械学習によるアノマリ検知システムの構築と回避 by Clarence Chio - CODE BLUE 2015
- 2. 機械学習
- 3. 私の目標 • 機械学習によるアノマリ検知システムの概略を伝える • いつ/どこで/どのように システムを構築するか議論に火をつける • システムが どれくらい 「安全」 か調査する • 私達はここからどこに向かうか議論する
- 6. 侵入検知？
- 8. 199.72.81.55 - - [01/Jul/1995:00:00:01 -0400] "GET /history/apollo/ HTTP/1.0" 200 6245 unicomp6.unicomp.net - - [01/Jul/1995:00:00:06 -0400] "GET /shuttle/countdown/ HTTP/1.0" 200 3985 199.120.110.21 - - [01/Jul/1995:00:00:09 -0400] "GET /shuttle/missions/sts-73/mission-sts- 73.html HTTP/1.0" 200 4085 burger.letters.com - - [01/Jul/1995:00:00:11 -0400] "GET /shuttle/countdown/liftoff.html HTTP/ 1.0" 304 0 199.120.110.21 - - [01/Jul/1995:00:00:11 -0400] "GET /shuttle/missions/sts-73/sts-73-patch- small.gif HTTP/1.0" 200 4179 burger.letters.com - - [01/Jul/1995:00:00:12 -0400] "GET /images/NASA-logosmall.gif HTTP/1.0" 304 0 burger.letters.com - - [01/Jul/1995:00:00:12 -0400] "GET /shuttle/countdown/video/ livevideo.gif HTTP/1.0" 200 0 205.212.115.106 - - [01/Jul/1995:00:00:12 -0400] "GET /shuttle/countdown/countdown.html HTTP/ 1.0" 200 3985 この中からどのようにアノマリを見つける？？？
- 9. 機械学習の技術は しきい値/ルールベースの ヒューリスティックと比べて なぜ魅力的なのか？ • 適応的 • 動的 • （理論的に）人の介入が最小
- 10. しかし…
- 11. しきい値/ルールベースの ヒューリスティックはなぜ良いか？ • 簡単な理由 • シンプル＆理解しやすい • 動的/適応的にも出来る
- 12. 好評な機械学習のアプリケーション
- 13. 好評な機械学習のアプリケーション
- 14. 好評な機械学習のアプリケーション
- 15. 期待値を設定
- 18. アノマリ検知は何が違うのか? 機械学習の他の問題点と基本的に異なる • とても多い誤検知 • 訓練データの不足 • 「意味上の隔たり」 • 評価の難しさ • 攻撃者の設定
- 24. どのようにして現実世界の アノマリ検知システムは 失敗してきたのか？ • 多くの偽陽性 • 攻撃がない訓練データの発見が大変 • 深く理解しないで使用 • ポイズニングモデル
- 25. それで・・・絶望した？
- 26. こうしよう！ • 時系列を生成する • 代表的な特徴を選び出す • 「正常」のモデルを訓練/調整する • 入力点がモデルから外れたら警告する
- 27. インフラの例 Sensitivity of PCA for Traffic Anomaly Detection, Ringberg et. al.
- 31. どのようにして特徴を選び出すか？ 難しさ： • 繰り返すべき組み合わせが非常に多い！ • 評価が大変 • 頻繁に「最適」が変わる • 基準だけではなく性能の精度 • 改良されたモデルの説明力 • より短い訓練期間 • 高度な一般化/削減された過学習
- 32. 主成分分析 • 自動的に特徴を選び出すための共通の 統計的な手法 どのように？ • データを異なる次元に変換する • データの分散を最もよく表すことの出来る 次元の順序付きのリスト（主成分）を返す
- 37. 「固有値」プロット 効果がより大きい 主成分分析 使用した主成分の数 累積的な%分散図 100804020060 10^0 10^1 10^2 10^3 効果がより小さい 主成分分析
- 38. どのように共通的な落とし穴を 回避するか？ • 脅威モデルをよく理解する • 検知範囲を狭めたままにする • 偽陰性/陽性の損失を減らす
- 39. 意味上の 隔たりを 補う
- 45. つまらないもの
- 48. 茹でガエル 攻撃
- 51. どのようにこれを防ぐか？ 後前
- 55. どのようにこれを防ぐか？ 改良された主成分分析 • 対策 • 主成分の追及 • ロバスト主成分分析
- 56. ロバスト統計 平均の代わりに中央値を使う 主成分分析の「分散」最大化 vs. 対策の「中央絶対偏差」 データセットをモデル化する適切な分布を見つける 正規/ガウス分布 vs. ラプラス分布 ロバスト主成分分析を使う
- 68. 間違った警告率（偽陽性率） ポイズニング検知率（真陽性率） 1.00.80.40.200.6 0 0.2 0.4 0.6 0.8 1.0 真陽性率 vs. 偽陽性率
- 69. 攻撃期間 （訓練期間の数） 回避成功率（偽陰性率） 1.00.80.40.200.6 0 2 4 6 8 10 つまらないものを注入 0% 10% 20% 30% 40% 50% 茹でガエル向け 単純な注入向け 回避成功率
- 70. 単純な主成分分析 ロバスト（な）主成分分析 単純なつまらな いものの注入 （50%の注入、 たった一つの訓練期間） ～ 76% 回避成功 ～ 14% 回避成功 茹でガエルの 注入 (10の学習期間) ～ 87% 回避成功 ～ 38% 回避成功
- 71. • あまり良くないけど、改良されている・・・ • 純粋な機械学習のアノマリ検知システムは ずっと妥協されやすい • 特徴やしきい値を見つけるために機械学習を使い、 静的ルールを用いてアノマリ検知を流動的に行う 今日のアノマリ検知システム
- 72. 次はなに？ • 他の人が作ったアノマリ検知システムでさらにテストする • ポイズニング技術に対する他の防御 • 機械学習モデルの耐障害モードで実験する
Editor's Notes
- What anomaly detection can do for you what anomaly detection cannot do for you When is anomaly detection helpful? When is anomaly detection not helpful?
- If you came into this with a blank slate, and wanted to set up an anomaly detection system, how would you do it? (step-by-step) come up with an hypothetical, specific use case, walk through the steps. generating time-series extracting features tuning model parameters evaluating
- If you came into this with a blank slate, and wanted to set up an anomaly detection system, how would you do it? (step-by-step) come up with an hypothetical, specific use case, walk through the steps. generating time-series extracting features tuning model parameters evaluating
- briefly evaluate various ways, classification models, why PCA is helpful (extract features in high dimensionality spaces)
- briefly evaluate various ways, classification models, why PCA is helpful (extract features in high dimensionality spaces)
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- given that the nature of traffic is often changing and non-uniform, how do you consistently select features in an online fashion that keeps up with the seasonality of the data? it turns out that this is unarguably the hardest part of outlier detection, and has to be done, because in very few scenarios can you just train a static model once and the detector will work fine forever. thinking about it, if that is indeed the case, then you don’t have to deal with the complexity of ML anomaly detection techniques. you can use a simple rule/detector engine to alert when datapoints that does not fit the expected trend arrive. Explain PCA, demo of PCA, visualization of what PCA looks like
- how would you go about evaluating the efficacy of the system? finding the true FNR of the detector - hard? non-trivial problem, but it all boils down to FNR vs FPR define fnr and fPr does FNR/FPR change over time?
- how would you go about evaluating the efficacy of the system? finding the true FNR of the detector - hard? non-trivial problem, but it all boils down to FNR vs FPR define fnr and fPr does FNR/FPR change over time?
- lets say you wanted to get past an anomaly detector we can make several different assumptions about the power that you have. what do you know? global knowledge? local knowledge?
- lets say you wanted to get past an anomaly detector we can make several different assumptions about the power that you have. what do you know? global knowledge? local knowledge?