ImageProof: Enabling Authentication for Large-Scale Image Retrieval Shangwei Guo 1 Jianliang Xu 1 Ce Zhang 1 Cheng Xu 1 Tao Xiang 2 1 Department of Computer Science, Hong Kong Baptist University 2 College of Computer Science, Chongqing University {csswguo,xujl,cezhang,chengxu}@comp.hkbu.edu.hk, txiang@cqu.edu.cn ICDE 2019 Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval
Background • Content-based image retrieval (CBIR) has been widely used in business • Data-as-a-Service (DaaS) enables companies to build and then outsource image retrieval systems to cloud platforms Service Provider Database Client Image Owner Similar images Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 1/17
Background • Content-based image retrieval (CBIR) has been widely used in business • Data-as-a-Service (DaaS) enables companies to build and then outsource image retrieval systems to cloud platforms Service Provider Database Client Image Owner Similar images • Security Threat : • Query result integrity not guaranteed due to software/hardware malfunctions, hack attacks • Examples • Product image search • Medical image search Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 1/17
SIFT-Based Image Retrieval • Detect and extract local features using scale invariant feature transform (SIFT) and its variants Feature Extraction Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 2/17
SIFT-Based Image Retrieval • Detect and extract local features using scale invariant feature transform (SIFT) and its variants • Twe Steps • Bag-of-visual-words (BoVW) encoding • Approximate k -means (AKM) using randomized k-d trees Feature Extraction BoVW Encoding AKM Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 2/17
SIFT-Based Image Retrieval • Detect and extract local features using scale invariant feature transform (SIFT) and its variants • Twe Steps • Bag-of-visual-words (BoVW) encoding • Approximate k -means (AKM) using randomized k-d trees • Inverted index search: search similar images with impact-ordered inverted index Feature Extraction BoVW Encoding AKM Inverted Index Search Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 2/17
Problem Model • Malicious threat model • The service provider (SP) could return incorrect results (e.g., faked or low- ranked images) Service Provider Database Client Image Owner Similar images Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 3/17
Problem Model • Malicious threat model • The service provider (SP) could return incorrect results (e.g., faked or low- ranked images) Service Provider Database Client Image Owner Similar images • Query authentication for SIFT-based image retrieval and top- k query Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 3/17
Problem Model • Malicious threat model • The service provider (SP) could return incorrect results (e.g., faked or low- ranked images) Service Provider Database Client Image Owner Similar images • Query authentication for SIFT-based image retrieval and top- k query • Challenges • Designing a query authentication scheme for a large and complex retrieval system is a big challenge in itself • The client usually has only limited storage, communication, and computation resources Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 3/17
Problem Model Service Provider Database and ADS Client Image Owner Similar images & VO • Our Solution : • Taking the advantage of the authenticated data structures (ADSs), the SP returns a verification object (VO) to prove • Soundness: The results must be the images which have not been tampered with • Completeness: The results include the k most similar images Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 4/17
Our Contributions • Propose an efficient authentication scheme, ImageProof, for SIFT-based image retrieval with large or medium-sized codebooks Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 5/17
Our Contributions • Propose an efficient authentication scheme, ImageProof, for SIFT-based image retrieval with large or medium-sized codebooks • Two novel ADS components: • Merkle randomized k-d tree • Merkle inverted index with cuckoo filters Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 5/17
Our Contributions • Propose an efficient authentication scheme, ImageProof, for SIFT-based image retrieval with large or medium-sized codebooks • Two novel ADS components: • Merkle randomized k-d tree • Merkle inverted index with cuckoo filters • Develop several optimization techniques to further reduce the costs of both the SP and the client Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 5/17
1 2 fp x 3 h 2 ( x ) 4 x 5 6 h 1 ( x ) fp z 7 Preliminaries • Merkle Hash Tree • An authenticated binary tree, enabling users to verify individual data objects without retrieving the entire database si g mht = sign ( sk , h N 1 ) N 1 h N 1 = h ( h N 2 h N 3 | ) N 2 N 3 h N 2 = h ( h N 4 h N 5 | ) N 4 N 5 N 6 N 7 h N 7 = h ( h ( o 7 ) | h ( o 8 )) o 1 o 2 o 3 o 4 o 5 o 6 o 7 o 8 Figure 1: An example of a Merkle hash tree. Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 6/17
Preliminaries • Merkle Hash Tree • An authenticated binary tree, enabling users to verify individual data objects without retrieving the entire database • Cuckoo Filter • An efficient data structure for approximate set membership tests • Two hash values per item • Support delete operation 1 si g mht = sign ( sk , h N 1 ) 2 N 1 fp x h N 1 = h ( h N 2 h N 3 | ) 3 h 2 ( x ) Insert Delete N 2 N 3 h N 2 = h ( h N 4 h N 5 | ) 4 x 5 N 4 N 5 N 6 N 7 h N 7 = h ( h ( o 7 ) | h ( o 8 )) 6 h 1 ( x ) fp z o 1 o 2 o 3 o 4 o 5 o 6 o 7 o 8 7 Figure 2: A cuckoo filter, two hash values Figure 1: An example of a Merkle hash per item. tree. Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 6/17
Scheme Overview • Ensure the integrity of query processing for each step Authenticated BoVW Encoding Authenticated Inverted Index Search Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 7/17
Scheme Overview • Ensure the integrity of query processing for each step • Two novel ADS components: • Merkle randomized k-d tree • Merkle inverted index with cuckoo filters Merkle Randomized k-d Tree Merkle Inverted Index Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 7/17
Merkle Randomlized k-d Tree (MRKD-tree) • ADS • Internal nodes and leaf nodes l 5 l 1 l 7 l 1 c 8 h 1 = h ( | l 1 h 2 h 3 | ) c 3 c 7 l 2 l 3 l 3 c 4 q 2 l 2 c 6 c 2 l 4 l 5 l 6 l 7 h 7 = h ( | l 7 h o 7 h o 8 | ) c 5 c 1 o 1 o 2 o 3 o 4 o 5 o 6 o 7 o 8 q 1 o 8 = { c 8 h Γ c 8 , }, h o 8 = h ( c 8 h Γ c 8 | ) l 4 l 6 l 1 h 2 l 3 l 6 h Γ c 5 h Γ c 6 , [ [ , ], h 7 The V O C for q 1 q 2 , : [ ]] Figure 3: An example of the MRKD-tree and VO generation for query q 1 , q 2 . Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 8/17
Merkle Randomlized k-d Tree (MRKD-tree) • ADS • Internal nodes and leaf nodes l 5 l 1 l 7 l 1 c 8 h 1 = h ( | l 1 h 2 h 3 | ) c 3 c 7 l 2 l 3 l 3 c 4 q 2 l 2 c 6 c 2 l 4 l 5 l 6 l 7 h 7 = h ( | l 7 h o 7 h o 8 | ) c 5 c 1 o 1 o 2 o 3 o 4 o 5 o 6 o 7 o 8 q 1 o 8 = { c 8 h Γ c 8 , }, h o 8 = h ( c 8 h Γ c 8 | ) l 4 l 6 l 1 h 2 l 3 l 6 h Γ c 5 h Γ c 6 , [ [ , ], h 7 The V O C for q 1 q 2 , : [ ]] Figure 3: An example of the MRKD-tree and VO generation for query q 1 , q 2 . • Authenticated Query Processing • Given a set of feature vectors, calculate the BoVW vector • Generate a single verification object (VO) for all feature vectors by maximiz- ing the use of shared tree nodes Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 8/17
Merkle Inverted Index With Cuckoo Filters • ADS • Each Merkle inverted list Γ c i consists of five components, i.e., the associated cluster c i , the digest h (Θ c i ) , the cluster weight w c i , the cuckoo filter Θ i and its posting list Table 1: An example of the Merkle inverted lists. Θ i c i h Γ ci w c i Posting Lists √ √ c 5 h ( 2 2 | h (Θ c 5 ) | h pos 5 , 1 ) 2 2 Θ c 5 �→ � 1 , 0 . 34 , h pos 5 , 1 � � 3 , 0 . 26 , h pos 5 , 2 � � 4 , 0 . 25 , h pos 5 , 3 � ... √ √ h ( 2 | h (Θ c 6 ) | h pos 6 , 1 ) Θ c 6 �→ � 5 , 0 . 41 , h pos 6 , 1 � � 8 , 0 . 32 , h pos 6 , 2 � � 3 , 0 . 28 , h pos 6 , 3 � c 6 2 ... Guo et al. | ImageProof: Enabling Authentication for Large-Scale Image Retrieval 9/17
Recommend
More recommend
