Towards Secure Edge-assisted Image Sharing for Timely Disaster

Situation Awareness

Jing Yao

1,2

, Yifeng Zheng

, Cong Wang

1,2

and Surya Nepal

City University of Hong Kong, Hong Kong, China

City University of Hong Kong, Shenzhen Research Institute, Shenzhen, China

CSIRO Data61 & CSCRC, Australia

Keywords:

Edge Computing, Data Security, Image Sharing.

Abstract:

To save human lives and reduce injury and property loss in disasters, it is important to collect real-time situa-

tion awareness information such as the surroundings, road conditions, resource information, and more. Among

others, images carry rich information and can easily provide a comprehensive view of the disaster situations.

This is nowadays greatly facilitated with the prevalence of camera-embedded smartphones. However, high

redundancy typically exists among the images gathered from different users during disasters. Given that band-

width is dearer in disaster situations, it would be valuable to detect the image redundancy during transmission

so that bandwidth allocation can be prioritized for unique images, enabling the timely delivery of useful in-

formation. In light of the above, in this position paper, we propose the design of an image sharing system

architecture for timely disaster situation awareness. Our system architecture takes advantage of the emerging

edge computing paradigm to perform image redundancy detection and prioritize the transmission of unique

images, optimizing the amount of useful information delivered within a certain period of time. Meanwhile, to

prevent images from being exposed to the intermediate edge infrastructure, our protocol is devised in a manner

that the edge infrastructure can perform image redundancy detection without seeing the images in the clear.

1 INTRODUCTION

Rapid relief is of paramount importance to save hu-

man lives and reduce injury and property loss in dis-

asters (earthquakes, ﬁres, tornadoes, etc). To facilitate

the rescue workers to effectively and rapidly perform

the work, it is essential to collect real-time situation

awareness information such as the surroundings and

individuals, disaster events, road conditions, resource

information, and more (Zuo et al., 2019).

Among others, images are a carrier of rich in-

formation and could easily present a comprehen-

sive view and description of the disaster situations.

With the prevalence of camera-embedded and cost-

effective smartphones (Zheng et al., 2018), it is very

easy for smartphone users to report and share visual

information about their surroundings through photos.

Indeed, it is quite common for users to take photos

and share them to help reﬂect the situation of disas-

ters. For example, in Typhoon Haiyan (2013), a huge

amount of images had been shared by users, which

had been leveraged by volunteers to explore which

place is in the greatest need of help (Zuo et al., 2019).

While the images shared during disasters provide

valuable information for timely situation awareness,

there usually exists large redundancy among the im-

ages uploaded by different users. Here, the redun-

dancy refers to images that capture observations on

the same objects/events. For example, a research

study (Weinsberg et al., 2012) shows that 22% of the

images taken by users during the Haiti earthquake

(2010) are similar to each other, and the percentage

even reaches 52% in the image set regarding the San

Diego ﬁre (2007).

Such high redundancy poses challenges on the

timely delivery of useful information for situation

awareness in disaster. In particular, if all images are

treated equally and directly transferred to the disaster

management service (DMS), it poses an obstacle for

gaining unique information in real-time. A lot of re-

sources and efforts would have to be put into the dis-

covery of unique images. Meanwhile, the transfer of

all images at the same time also imposes a heavy load

on the network, which would be highly challenging

given that network bandwidth is dearer in disasters. It

would be much more practical to prioritize the band-

Yao, J., Zheng, Y., Wang, C. and Nepal, S.

Towards Secure Edge-assisted Image Sharing for Timely Disaster Situation Awareness.

DOI: 10.5220/0009801602950301

In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - SECRYPT, pages 295-301

ISBN: 978-989-758-446-6

 2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

295

width allocation on unique images. Therefore, how

to properly prioritize the upload of images from dif-

ferent users for the DMS in disaster scenarios is of

critical importance for timely situation awareness.

To support this, one promising direction is to take

advantage of edge computing, an emerging paradigm

which extends the capabilities of cloud computing

to the network edge (NetworkWorld, 2017). Among

others, one notable advantage of edge computing for

disaster response is that it would allow routers, smart-

phones, and other devices to keep collecting data

even without Internet connections (StorageCraft.com,

2020). So, with edge computing deployed between

the cloud-based DMS and the users, this opens up

the opportunities of giving fast responses to users and

performing redundancy detection before images are

uploaded to the DMS. As the edge infrastructure and

the DMS could be in different trust domains and the

service is set up by the DMS, a crucial requirement

here is that the processing at the intermediate edge

infrastructure should be done over images in the ci-

phertext domain, which prevents third-parties from

accessing/learning about the images.

In this position paper, we propose the design of

a system architecture enabling secure edge-assisted

image sharing for the DMS to gain timely situation

awareness in disasters. At a high level, our architec-

ture takes advantage of the emerging edge computing

architecture to perform image redundancy detection

and prioritize the transmission of the unique images

to the DMS, so as to optimize the amount of useful in-

formation from image sharing within a certain period

of time. For redundancy detection, we mainly rely

on global features extracted from images and lever-

age them to perform similarity measurement between

different images. To support secure redundancy de-

tection efﬁciently, we resort to an effective similar-

ity search technique called locality-sensitive hashing

(LSH) so that the problem of similarity measurement

of encrypted images can be transformed into equality

testing over the protected image features.

With this as a basis, we then propose to perform

secure in-batch redundancy detection and cross-batch

redundancy detection, with the goal of minimizing the

redundancy per image delivery from the edge server

to the DMS. Here, protected in-batch redundancy de-

tection allows the edge server to detect image redun-

dancy across a batch of images uploaded from dif-

ferent users with a certain period of time and out-

put some candidates of representative images. And

secure cross-batch redundancy detection allows the

edge server to further detect image redundancy be-

tween the representative images and the images previ-

ously uploaded to the DMS via the edge server. With

Edge Server

Mobile Client

Edge Server

Mobile Client

Disaster Management Service

Figure 1: Overview of the system architecture.

the synergy of these protected redundancy detection

strategies, it could provide as much useful and unique

information as possible per image delivery from the

edge server to the DMS.

The rest of this position paper is organized as fol-

lows. Section 2 presents our problem statement. Sec-

tion 3 introduces some preliminaries. Section 4 gives

the details of our design. Section 5 describes the re-

lated work. Section 6 makes conclusions and indi-

cates the future work for this position paper.

2 PROBLEM STATEMENT

2.1 System Architecture

Fig. 1 shows the envisioned system architecture of

secure edge-assisted image sharing for timely situa-

tional awareness in disasters. There are three types

of actors: the mobile client, the cloud-based disas-

ter management service (DMS), and the edge server.

The mobile client is a software application which en-

ables one to share photos captured about disaster situ-

ations to the DMS. The DMS could be setup by re-

lated organization such as the emergency organiza-

tion, the local government, or a social media provider

(Nishiyama et al., 2017), which wants to gather rich

information rapidly so as to facilitate the arrangement

of the rescue activities and decision making for disas-

ter response. The edge server, which could be hosted

by an Internet Service Provider (ISP), is located at the

network edge and closer to the mobile clients. It facil-

itates the mobile clients to quickly share images dur-

ing disasters (NetworkWorld, 2017) without the need

to directly contact the remote DMS. The edge server

collects the images from the mobile clients in the dis-

aster area, and directly performs image redundancy

detection so as to optimize the amount of useful infor-

mation delivered from image sharing for timely situ-

ational awareness in disasters.

From practical considerations, it would be the

DMS’s desire to prevent the intermediate edge server

from directly accessing and learning about the im-

SECRYPT 2020 - 17th International Conference on Security and Cryptography

296

Moblie clients

Image capture

Edge server

Image feature index

DMS

Image storage

Image encryption

Image fingerprint

generation

Image fingerprint

encryption

Image redundancy detection

Enc. in-batch detection

Enc. cross-batch detection

Image transmission

Delay

transmission

Immediate

transmission

Figure 2: Service workﬂow.

ages, while still being able to perform the image re-

dundancy detection. This is because the whole service

is setup by the DMS, and the DMS and the edge server

could belong to different trust domains. Therefore,

we craft our design to support processing without dis-

closing the information of images. At a high level, as

shown in Fig. 2, the service ﬂow in our system is as

follows. In an initialization stage, each mobile client

signs up at DMS and obtains secret keys. In disas-

ter environments, to share images, each mobile client

ﬁrst generates ﬁngerprints for images , and then lever-

ages hash function to protect the ﬁngerprints. At last,

each mobile client encrypts images. The protected

ﬁngerprints and encrypted images are then sent to the

nearby edge server. Upon receiving a batch of pro-

tected ﬁngerprints and encrypted images from possi-

bly multiple clients within a certain period of time, the

edge server will prioritize the image transmission so

as to decide which images should be uploaded with

high priority. Speciﬁcally, for this batch of images,

the edge server leverages the protected ﬁngerprints to

do redundancy (similarity) detection, so as to select

a subset of representative encrypted images. We call

this procedure secure in-batch redundancy detection.

For these representative encrypted images, the

edge server then performs similarity detection based

on their protected ﬁngerprints, against the set A of

protected ﬁngerprints of images previously uploaded

via the edge server. We call this secure cross-batch

similarity detection. For those non-similar encrypted

images, the edge server places them ino a set B, and

also adds their protected ﬁngerprints to the set A. Af-

ter secure cross-batch similarity detection, the edge

server puts the images in B into the transmission

queue, in the order of their arrival, and initiates the

image transmission to the DMS. For the remaining

images, the edge server puts them into another queue

pending transmission, in the order of their arrival.

2.2 Threat Model

We consider that the threats in our system architecture

mainly come from the engagement of the edge server.

Our security goal is to protect the conﬁdentiality of

the images as well as the ﬁngerprints, against the in-

termediate edge server. It is assumed to be honest-

but-curious, which means that it will follow our pro-

tocol faithfully yet attempt to obtain the images. We

note that this assumption is widely adopted in the lit-

erature so our adoption is consistent with prior work

(Ma et al., 2019). Note that the ﬁngerprint of an im-

age could be used to infer some information of that

image (Ferreira et al., 2017), so all the ﬁngerprints

also demand protection. Additionally, we assume that

the mobile client and the DMS are fully trusted. We

deem attacks like DDoS attacks and data integrity at-

tacks out of the scope.

3 PRELIMINARIES

3.1 Image Features

Image features could provide a short summary of the

content of an image, i.e., acting as a ﬁngerprint. In

general, image features can be divided into two cate-

gories: (1) local features and (2) global features. A

local feature only describes an interesting point of

an image. So, for an image, multiple local features

need to be computed. Different from local features, a

global feature can represent the whole content of an

image. It is usually computed by the color histogram,

texture values of an image, etc. In comparison with

local features, global features are advantageous due

to their high efﬁciency in feature extraction as well as

in use for similarity detection (Chamoso et al., 2018).

Therefore, we use global features to detect image sim-

ilarity.

3.2 Locality-sensitive Hashing

A locality-sensitive hashing distributes a hash func-

tion family over a dataset, enabling similar data points

to have hash collisions with high probability and dis-

similar data to have hash collisions with little proba-

bility. It is generally used for similarity search(Zheng

et al., 2017; Cui et al., 2016), with the formal deﬁni-

tion stated as follows.

Deﬁnition 1. (Locality-sensitive Hashing) Let F be a

hash function family and R be the range of hash re-

sults of the dataset D. The distance function is de-

noted as dis(x, y), where x, y ∈ D. Given the two

distance values {(r

, r

)|r

< r

} and the two prob-

ability values {(p

, p

)|p

> p

}, the family F =

{h : D → R} is a (r

, r

, p

)-locality sensitive

hash function family if for any {(d

, d

)|dis(d

, d

) ≤

∧ d

, d

∈ D}, then Pr[h(d

) = h(d

)] ≥ p

; if

Towards Secure Edge-assisted Image Sharing for Timely Disaster Situation Awareness

297

for any {(d

, d

)|dis(d

, d

) > r

∧ d

, d

∈ D}, then

Pr[h(d

) = h(d

)] ≤ p

4 DESIGN OF SECURE

EDGE-ASSISTED IMAGE

SHARING IN DISASTERS

In this section, we start with the description of our de-

sign intuition on secure edge-assisted image sharing

in disasters. Then, we move on to the introduction of

the detailed scheme.

4.1 Design Intuition

Our main design insight is to prioritize the transmis-

sion of the unique images via ﬁngerprint-based sim-

ilarity detection at the edge server, so as to optimize

the amount of useful information per image delivery

for timely situational awareness in disasters. As we

aim to protect the conﬁdentiality of images and ﬁn-

gerprints, all the processing should be done with the

prerequisite of protecting the conﬁdentiality of im-

ages and ﬁngerprints. To instantiate our main idea,

some design considerations need to be addressed.

The ﬁrst consideration is how to efﬁciently sup-

port similarity detection for two encrypted images.

Recall that we resort to global features as ﬁngerprints

for similarity detection. The straightforward way of

doing this is to compute the distance (under a com-

mon metric like Hamming distance or Euclidean dis-

tance) between the ﬁngerprints of two images. A

smaller distance means a higher similarity between

the two images. If the distance is less than a pre-

deﬁned threshold, the two images are considered as

similar images. However, such a plausible approach

would demand the use of advanced and expensive

cryptographic primitives like homomorphic encryp-

tion in our scenario. To avoid pairwise distance com-

putation between the ﬁngerprints as well as achieve

high efﬁciency in similarity detection, we leverage a

set of LSH functions to hash global features of im-

ages. According to the deﬁnition of LSH, an LSH

function can map the features of similar images to

identical LSH values with high probability. In this

case, the similarity of any two images can be ef-

ﬁciently measured by counting the number of their

matched LSH values.

The second consideration is on the conﬁdential-

ity of the image ﬁngerprints. As LSH does not have

the one-way property from a cryptographic perspec-

tive (Partridge et al., 2012), the adversary might still

be able to infer information about the image ﬁnger-

prints. To preserve the conﬁdentiality of the ﬁnger-

prints, our design hashes LSH values before the mo-

bile client’s uploading them to the edge server. Con-

sidering the resource constraints of mobile devices,

we take advantage of a lightweight mechanism, i.e.,

one-way hash, for mobile clients to protect the ﬁnger-

prints. In order to preserve the characteristic of LSH

that maps similar data to the identical hash result with

high probability, each LSH value of the ﬁngerprint

is hashed separately. Note that the similarity detec-

tion is only based on the ﬁngerprints so the images

could be independently encrypted under standard en-

cryption mechanisms.

The third consideration is how to properly perform

secure in-batch redundancy detection and cross-batch

redundancy detection. As mentioned before, secure

in-batch redundancy detection refers to select a sub-

set of representative encrypted images from a batch

of encrypted images collected by the edge server over

a certain period of time. In contrast, secure cross-

batch redundancy detection aims to select images

from the representative image set so that the selected

images ware not similar to the images previously up-

loaded by the edge server. It is noted that most of

the existing designs only consider the detection of

cross-batch redundancy, such as SmartEye(Hua et al.,

2015) and MRC(Dao et al., 2017). The only design

that takes into account both cross-batch redundancy

and in-batch redundancy is the design due to Zuo et

al. (Zuo et al., 2019). Their design works under a

different scenario where only the cloud is engaged.

Besides, they ﬁrst detect the cross-batch similar im-

ages, followed by the detection of the in-batch similar

images. Given that the similar images will be trans-

mitted in the same time period with high probability

under the disaster scenario, our belief is that it would

be much more efﬁcient to ﬁrst perform in-batch simi-

larity detection and then cross-batch similarity detec-

tion. Through this delicate consideration, the number

of images involved in cross-batch similarity detection

would be largely reduced, leading to efﬁciency im-

provement on the cost-dominant cross-batch redun-

dancy detection.

4.2 Our Proposed Design

We now present the detailed design for secure edge-

assisted image sharing in disasters. Overall, there are

three key modules, i.e., data preparation, secure in-

batch redundancy detection, and secure cross-batch

redundancy detection. We elaborate on each of them

as follows.

SECRYPT 2020 - 17th International Conference on Security and Cryptography

298

Algorithm 1: Data Preparation at the Mobile Client.

Input: Image I and secret key k.

Output: Encrypted image c and protected feature v.

1: f ← f Extract (I).

2: w ← LSH ( f ), where w =

{

, · · · , w

}

;

3: for i ∈

{

1, · · · , l

}

4: v

← H (w

5: end for

6: c ← E

(Im).

4.2.1 Data Preparation

Fig. 1 shows the data preparation module. The mo-

bile client ﬁrst calls the feature extraction function

f Extract (·) to extract the ﬁngerprint f from the im-

age Im to be uploaded. Examples of global features as

ﬁngerprints include dhash and ahash(Neal Krawetz,

2013). The mobile client then applies l LSH func-

tions to the ﬁngerprint and generates a set w of LSH

values. At last, the mobile client hashes w with a one-

way hash function H (·) and outputs the protected ﬁn-

gerprint v =

{

, · · · , v

}

. For the image I, the mobile

client calls an encryption scheme E (·) which could be

the standard AES encryption scheme, and outputs the

encrypted image c. Finally, the mobile client uploads

all encrypted images and protected ﬁngerprints to the

nearby edge server.

4.2.2 Secure In-batch Redundancy Detection

In disaster scenarios, it is very likely that there are

similar images in the batch of images collected the

edge server over a certain period of time. Therefore,

we perform in-batch redundancy detection so as to se-

lect a representative image set from these images. The

high-level idea is to ﬁrst divide the batch of images

into groups according to the similarity between each

of two encrypted images in the batch. In particular,

the edge server ﬁrst calculates the similarity between

each of any two images in the batch. Let the protected

ﬁngerprints of two images I

and I

be v

and v

re-

spectively. The distance between the two images is

computed via dis (I

, I

) = ||v

− v

, where || · ||

denotes the L1 distance. The two images I

and I

are considered to be similar, if dis (I

, I

) < ε, where

ε is a pre-deﬁned similarity threshold. Based on this,

the edge server then divides the in-batch images into

groups. We use k to denote the number of groups.

According to the grouping result, the edge server

then proceeds to select a set of representative images

in the batch. There are several intuitive strategies that

could be adopted here. The ﬁrst strategy is to select

from each group the image that arrives at the edge

server the earliest, following the general principles of

packet forwarding for routers and switches, i.e., ﬁrst-

in, ﬁrst-out (FIFO). However, this strategy may lead

to an excessive content loss, given that it is very likely

that the content of the ﬁrst image may not well sum-

marize the information contained in this group of im-

ages. The second possible strategy is to select the im-

age with the highest resolution in each group. How-

ever, this also does not guarantee that the largest im-

age is the most informative one in the group.

In short, although these intuitive strategies are

simple and relatively easy to realize, they may not

produce a good representative set that could well

represent the whole batch of images. We note that

an alternative choice is to adapt a more advanced

optimization-based strategy inspired by (Zuo et al.,

2019) in our scenario, which allows to select a sub-

set T of images that could best represent the whole

batch S of images. At a high level, the selection

of a subset of representative images is formulated as

an optimization problem, where the objective is for-

mulated as: T

∗

∈ argmax

T ⊂S

F(T ). Here, F is a

scoring function that can quantitatively represent the

quality of a summary (i.e., the subset of representa-

tive images). The realization of the function F is a

weighted sum of two sub-functions: a coverage func-

tion f

and diversity function f

. That is, F(T ) =

· f

(T ) + λ

· f

(T ), where λ

and λ

are non-

negative. Speciﬁcally, the coverage function is for-

mulated as f

∑

i∈S

max

j∈T

i, j

, where the intuition

is to use the sum of the similarity between an image

i in S and its most similar image in T . For the diver-

sity function, it is formulated as f

∑

i=1

N (T, I

where I

refers to the set of images in a i-th group,

and N (T, I

) takes the value 0 if T and I

have empty

intersection, and the value 1 otherwise. Solving the

above optimization needs a constraint on the size of

the selected subset, for which we set it to the number

of groups, i.e., |T | ≤ k.

Algorithm 2 shows the procedure of in-batch re-

dundancy detection, which allows the edge server to

select a representative subset of images from a batch

of images collected over a certain period of time. The

edge server ﬁrst calculates the similarity between any

two images in the batch. Then, it uses the similar-

ity threshold ε to partition the batch of images into

several groups. Finally, the edge server solves the op-

timization problem via a greedy algorithm to output a

subset of representative images.

4.2.3 Secure Cross-batch Redundancy Detection

In secure cross-batch redundancy detection, we aim

to further select a subset from the encrypted represen-

tative images obtained from in-batch similarity detec-

Towards Secure Edge-assisted Image Sharing for Timely Disaster Situation Awareness

299

Algorithm 2: Secure In-batch Redundancy Detection at the

Edge Server.

Input: The batch S of encrypted images C

, · · · , c

|S|

} and the batch of protected features

= {v

, · · · , v

|S|

Output: T

(n is the number of iterations).

1: Compute pairwise similarity w

i, j

among the im-

ages C

in the batch S according to the batch of

protected features V

2: Divide the batch S of images into several groups

according to the principle of whether w

i, j

< ε.

3: Denote the number of groups as k.

4: Choose an image S

arbitrarily.

5: T

← S

6: While |T

| ≤ k

7: Choose S

∈ argmax

∈S\T

F(T

∪ {S

}).

8: T

i+1

← T

∪ {S

9: i ← i + 1.

10: end while

tion, ensuring that each of the selected images is not

similar to any image uploaded to the DMS before via

the edge server. The high-level idea of secure cross-

batch redundancy detection is to query the protected

ﬁngerprints of the encrypted images previously up-

loaded to the DMS, and check if some encrypted rep-

resentative images are similar to them.

A simple approach for the edge server is to send

the protected ﬁngerprints of all the representative im-

ages to the DMS, which then performs the redun-

dancy detection and returns the detection result to the

edge server. However, this could lead to high latency.

To alleviate this problem, our design is to let the edge

server build and maintain an index using efﬁcient and

succinct data structures over the protected ﬁngerprints

of images previously uploaded to the DMS. In this

case, the edge server can perform cross-batch redun-

dancy detection locally. When the edge server sends

an image to DMS, it also inserts the protected ﬁnger-

print of that image to the index.

4.3 Security Guarantees

The security of our proposed design directly follows

from that of the one-way hash and the image encryp-

tion scheme. Speciﬁcally, as we use a one-way hash

function to hash the LSH values of image ﬁngerprints,

this makes it computationally infeasible to reveal the

ﬁngerprint from the protected ﬁngerprints (Partridge

et al., 2012). Therefore, the proposed design can pro-

tect the image ﬁngerprints against the edge server. For

the images, recall they are encrypted via either stan-

dard encryption schemes. So the conﬁdentiality of the

images is well protected against the edger server. Ad-

ditionally, we remark that revealing the similarity of

images in the ciphertext domain is necessary to sup-

port the target functionality of redundancy detection.

5 RELATED WORK

Our work is closely related to the existing works on

content-aware redundancy elimination in disaster en-

vironments, where the ultimate goal is to ask mobile

clients to only upload unique images after image re-

dundancy detection. Some works ((Dao et al., 2017;

Zuo et al., 2019), to just list a few) have proposed

to perform redundancy detection at the remote server

side (e.g., the cloud). In these designs, the mobile

client is typically required to send all features and

wait for the detection result. This might greatly re-

duce user experience when the database of images at

the remote server side is very large, which could be

the norm in disaster scenarios.

Instead of doing redundancy detection at the re-

mote server side, some works ((Hua et al., 2015; Zuo

et al., 2019; Weinsberg et al., 2012), to just list a

few) have proposed to take advantage of advanced

networking architectures like software-deﬁned net-

working (SDN) or delay tolerant network (DTN) to

perform in-network redundancy detection. Our work

differs from prior work mainly from several aspects.

Firstly, our design takes into account the security of

images and ﬁngerprints so the redundancy detection

is all performed while protecting the conﬁdentiality of

images. Secondly, our design takes advantage of the

emerging edge computing architecture to perform se-

cure redundancy detection. With the assistance of the

edge server, the mobile clients in our system archi-

tecture do not need to wait for the redundancy detec-

tion results. Thirdly, while most of the existing works

only consider cross-batch redundancy detection, our

design support both (secure) in-batch redundancy de-

tection and cross-batch redundancy detection, which

is customized for our scenario as well.

6 CONCLUSIONS AND FUTURE

WORK

In this position paper, we explore and propose a sys-

tem architecture for the DMS, enabling secure edge-

assisted image sharing for timely disaster awareness.

Aiming for redundancy detection as well as image se-

curity against the intermediate edge infrastructure, we

propose to apply the strategies of in-batch redundancy

SECRYPT 2020 - 17th International Conference on Security and Cryptography

300

detection and cross-batch redundancy detection, and

design from the ground up so as to do all the effec-

tive processing while ensuring protection of images.

Our customized design comes from a synergy of a se-

ries of techniques including image processing, data

encryption, efﬁcient similarity search, and optimiza-

tion.

As future work, we plan to implement a proof-

of-concept system prototype and conduct a compre-

hensive evaluation over real-world datasets. Specif-

ically, we will measure the effectiveness of our se-

cure redundancy detection design over some real-

world disaster image datasets. We will also evalu-

ate the cost efﬁciency on different ends along the ser-

vice ﬂow. We also intend to deﬁne formal security

deﬁnitions and provide formal proofs. Besides, we

will explore the design space of emerging security so-

lutions like trusted execution environments, for efﬁ-

ciently defending against malicious adversaries that

compromise the edge server and deviate arbitrarily.

ACKNOWLEDGEMENTS

This work was supported in part by the Research

Grants Council of Hong Kong under Grants CityU

11202419, CityU 11212717, CityU 11217819, and

CityU C1008-16G.

REFERENCES

Chamoso, P., Rivas, A., S

anchez-Torres, R., and Rodr

ıguez,

S. (2018). Social computing for image matching. PloS

one, 13(5):e0197576.

Cui, H., Yuan, X., Zheng, Y., and Wang, C. (2016). En-

abling secure and effective near-duplicate detection

over encrypted in-network storage. In Proc. of IEEE

INFOCOM.

Dao, T., Roy-Chowdhury, A. K., Madhyastha, H. V., Krish-

namurthy, S. V., and Porta, T. L. (2017). Managing

redundant content in bandwidth constrained wireless

networks. IEEE/ACM TON, 25(2):988–1003.

Ferreira, B., Rodrigues, J., Leitao, J., and Domingos, H.

(2017). Practical privacy-preserving content-based

retrieval in cloud image repositories. IEEE TCC,

13(9):1–14.

Hua, Y., He, W., Liu, X., and Feng, D. (2015). Smarteye:

Real-time and efﬁcient cloud image sharing for disas-

ter environments. In Proc. of IEEE INFOCOM.

Ma, L., Liu, X., Pei, Q., and Xiang, Y. (2019). Privacy-

preserving reputation management for edge comput-

ing enhanced mobile crowdsensing. IEEE Trans. Ser-

vices Computing, 12(5):786–799.

Neal Krawetz (2013). Kind of Like That.

http://www.hackerfactor.com/blog/?/archives/

529-Kind-of-Like-That.html/.

NetworkWorld (2017). When disasters strike,

edge computing must kick in. https:

//www.networkworld.com/article/3228884/

when-disasters-strike-edge-computing-must-kick-in.

html.

Nishiyama, J., Tabata, S., and Shigeno, H. (2017). An efﬁ-

cient image gathering scheme with quality control in

disaster. In Proc. of IEEE AINA.

Partridge, K., Pathak, M. A., Uzun, E., and Wang, C.

(2012). Picoda: Privacy-preserving smart coupon de-

livery architecture. In Proc. of HotPETs.

StorageCraft.com (2020). Edge Computing vs.

Cloud Computing. https://blog.storagecraft.com/

edge-computing-cloud-computing/.

Weinsberg, U., Li, Q., Taft, N., Balachandran, A., Sekar,

V., Iannaccone, G., and Seshan, S. (2012). CARE:

content aware redundancy elimination for challenged

networks. In Proc. of ACM HotNets.

Zheng, Y., Cui, H., Wang, C., and Zhou, J. (2017).

Privacy-preserving image denoising from external

cloud databases. IEEE Trans. Information Forensics

and Security, 12(6):1285–1298.

Zheng, Y., Duan, H., and Wang, C. (2018). Learn-

ing the truth privately and conﬁdently: Encrypted

conﬁdence-aware truth discovery in mobile crowd-

sensing. IEEE Trans. Information Forensics and Se-

curity, 13(10):2475–2489.

Zuo, P., Hua, Y., Sun, Y., Liu, X. S., Wu, J., Guo, Y., Xia,

W., Cao, S., and Feng, D. (2019). Bandwidth and en-

ergy efﬁcient image sharing for situation awareness in

disasters. IEEE TPDS, 30(1):15–28.

Towards Secure Edge-assisted Image Sharing for Timely Disaster Situation Awareness

301