일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | ||||||
2 | 3 | 4 | 5 | 6 | 7 | 8 |
9 | 10 | 11 | 12 | 13 | 14 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 |
23 | 24 | 25 | 26 | 27 | 28 | 29 |
30 | 31 |
Tags
- neighboring pixel relationships
- ML
- linear inverse problem
- f2dnet
- semi-supervied learning
- Ai
- ddim
- conditional diffusion
- ilvr
- focal detection network
- ssda-yolo
- face forgery detection
- facenet
- unconditional generative models
- pedestrian detection
- hqs-algorithm
- fourmer
- diffpir
- deepfake detection
- image restoration
- unconditional ddpm
- IR
- daod
- diffusion
- ddpm
- Triplet
- object detection
- stochastic contraction theory
- ddrm
- ccdf
Archives
- Today
- Total
Stand on the shoulders of giants
[paper review] Triplet Fingerprinting 본문
Keys
-
NSL with triplet networks
- 새로 수집한 데이터 바로 모델학습에 사용가능/ 적은자원만으로도(weak adversary) WF 가능 [Threat landscape]
Abstract & introduction
- 기존것의 한계
- DF = 큰 데이터셋을 요구받음, 꾸준히 update해야함
- 대부분의 가정
- train/test data가 비슷한 distribution을 갖는다.
- train/test data가 “같은 시간”에 “같은 type”의 network condition에서 수집되었다.
- 공격자는 weak attacker model (realistic condition에 맞는)
Background
- Bootstrap time = 시스템 초기화 및 가동 과정이 완료되는 데 소요되는 시간
- (논문) ready-to-use classifier가될때까지 attacker에게 필요한 총시
- N-shot learning = NSL
- TLS connection = HTTPS를 암호화해주는 프로토
- protocol = 컴퓨터 내부에서, 또는 컴퓨터 사이에서 데이터의 교환 방식을 정의하는 규칙 체계
- ex. https, http(웹), TCP/IP (인터넷), SMTP(Simple Mail Transfer protocol, 이메일)
- gateway = 하나의 네트워크를 벗어나 다른 네트워크에 접속하는 장비를 게이트웨이라고 부른다. 무선 네트워크를 유선 네트워크에 연결시켜 주는 유무선 인터넷 공유기도 넓게 보면 게이트웨이이며 또한 라우터이다.
- stale ~= changing
- Keras는 API고 Tensorflow는 framework
WF Attack Assumption
1) Closed vs Open world Scenario
- size of open world
- unmonitored websites을 증가시키면서 모델 성능을 비교 평가했음
- 그런데 사실상 victim이 방문할 수있는 가짓수 >>> 수집량
- open-world evaluation model
- standard model = training data에 unmonitored set 포함 o
- AWF model = training data에 unmonitored set 포함 x,대신 cross-entropy loss 구해서 confidence threshold 기반으로 판정함.
2) Users’ Browsing Behaviour
- Tor clients가 specific behavior를 따름을 가정함
- browse sites sequentially
- use only single tab at a time
3) Traffic Parsing and Background traffic
- attacker은 site에서 만들어진 모든 traffic을 수집하고, 또 traffic들을 distinguish할 수 있다고 가정함.
- client와 guardnode ⇒ all Tor tracffic iz multiplexed in TLS connection
Attacker Goal
* 경시되었던 elements of WF attacks
- Generalizability
- attacker가 victim과 same condition에서 classifier을 학습시킴
- Bootstrap time
- assumption
- train, test → same distributioln
- real : time gap 존재 → mismatch → re-train 필요
- assumption
- Flexibility & Tranferability
- 일반적은 train, pred절차는 k monitored sites, T training instances for each monitored sites, U training instances for the unmonitored set이라고 할 때 k*T + U로 train하고 predict
- 문제점 = constraint 존재 → k개 websites만이 predicted class후보군임 → re-train 필연적
- 해결책 = N-shot learning, transfer learning
- Attack performance ⇒ 높은 성능은 유지
N-shot Learning(NSL) Technique
- 극소량의 데이터를 가지고 빠르게 모델을 학습시키자
- N examples per class
- 기존
- large dataset
- similar distribution in train, test
Triplet Network
- triplets (input) → 각각을 sub-network에 통과시킴.
- Anchor(A)
- Positive(P) : A와 같은 class
- Negative(N) : ≠ A의 class
- 이중에 하나만 가져오는데 그 선택 과정이나 근거나 하나만 가져오는 이유 이런거 설명..x (궁금점)
same website와의 distance(D)는 minimize, diff와는 maximize (~ contrastive learning)
Hyperparameter tunings
- base model
- distance metrics
- cosine distance (> Euclidean distance)
- mining strategy→ sol : Hard negative examples 사용
- strategies
- Random, Hard-Negative, Semi-Hard-Negative
- ( - ) 대부분의 negative examples N은 A와 too far
TF Attack Implementation
1. pre-training
fine-tuning hyperparmeters 및 sub-network으로 DF model사용
(softmax를 new FC layer로 대체)
2. Attack
2-1. N-training
- N-shot맞게 데이터 수집
- pre-training으로 얻은 모델(feature extractor)에 해당 샘플들 통과시켜 embedded vectors를 각 site마다 얻음
- 그 embedded vector들을 k-NN(SVM, MLP등)를 train시킴.
2-2. Classification
Experiments
- dataset
- Wang
- AWF
- DF
- timestamp & packet size 무시, direction only
- Similar but mutually exclusive datasets
- experimental settings
- N-ALL vs N-MEV (representation)
- N-ALL : k-NN을 N-train할때 N embedded samples사용
- N-MEV : k-NN을 N-train할때 mean of all N examples 시행(1
- ⇒ MEV(Mean Embedded Vector)
- results즉, 동일 class의 vector들을 average하다보니 embedded vector간의 noise를 reduce할 수 있었음.
- : N-MEV가 attack 정확도 더 높았음
- TF Goals
- overlaps btw pre-training and attack phases→ 오히려 overfitting & more rigid model
- → NSL with triplet networks 사용시 input들의 pair을 differentiate할 수 있어 효과적인 sample 학습이 가능함. (low traffic settings에서도 고성능, ↔ CUMUL, k-FP)
- → pre-training과 attack에 같은 class의 sample들을 포함하는것은 딱히 attack성능 향상에 도움이 안되었음.
- WF attacks with different data distributions
- experimental settings
- feature extraction = AWF775
- N-train = Wang100 → 둘이 아주 다름을 부록A에서 입증
- results
- 5-shot = 85%
- TF Goals
- feature extract dist ≠ attack
- experimental settings
- pre-training, classification datsets가 다른 시점, 다른 TBB에서 구축됨(→diff distributions)
- Transfer Learning vs Triplet Networks
- N total layers에서 앞의 k개 layer들을 freezing후 re-train the model
- → 마지막 FC layer 제외 모두 freeze할때 가장 성능이 좋았음.
- results
- case 1)N-shot에서 N이 작을땐 확실히 TF > Transfer learning(=Traditional)
- case 2) TF >> “
- 원인 분석
- Transfer learning은 fine-tuning할 수 있는 layer들이 별로 없다는 점..but
- TF는 feature extraction에서 모든 depth를 고려하니까.
- 원인 분석
- Open-world Scenario
- monitored / unmonitored sites 구분
- precision/ recall metrics 사용해서 평가
- experimental setting
- case1) mon=AWF100, unmon=AWF9000
- case2) mon=Wang100, unmon=AWF9000
- results
- ⇒ 두 settings에서 모두 effective (dist가 다를때 오히려 더 잘나왔음?!) ↔closed
- closed/open 결과의 비일관성의 원인
- = 많은 monitored sites가 unique content여서..(closed-world에서 하나하나 다 분류해내는게 mon/unmon 구분하는것보다 어려웠기때문이지 않을까)
- TF attack in the Larger Open World
- AWF400K = 현존 unmonitored 데이터셋중 가장 큰거
- Result = dataset이 커질수록 성능이 줄었음.
- TF attack against WTF-PAD
- experimental setting
- WTF-PAD로 defended traces simulate
- results
- 두 case모두 non-defended dataset일때보다 defended일때 attacks의 정확도급감 👎
- experimental setting
Appendix A: Dataset’s distribution analysis (B,C는 생략)
1. Basic network traffic statistic
- # of incoming/outcoming packets
- # of incoming/outcoming bursts
=> Wang & AWF dataset의 공통 sites에서 얘네들의 변화가 컸음.
2. Similarity measurement
- Cosine distance
- intrasimilarity = 내부 동일 website간
- intersimilarity = Wang과 AWF의 동일 website간
TF_TripletFingerprinting_3319535.3354217.pdf
2.39MB
Highlight version of the paper
기타
- WF자체가 웹사이트 마다 own unique pattern을 갖는다는 특징을 기반으로 만들어진 문제인데 distribution이 달라지는걸 건들여도 되는 이유
= (추측) distribution이 feature을 모두 설명해주지 못한다(pattern과 동치가 되지 못한다)
- 3년간 데이터를 사용하고 클래스당 5개 샘플 썼을 때 정확도 85% 달성함
-
NSL with triplet networks
- 장점: 새로수집한 데이터 바로 모델학습에 사용가능/ 적은자원만으로도(weak adversary) WF 가능[Threat landscape]
- 단점: protected traffic에 취약(defense 적용된 traces, ex. WTF-PAD)