Memory-Saving Oblivious RAM for Trajectory Data via Hierarchical

Generation of Dummy Access over Untrusted Cloud Environment

Taisho Sasada

1 a

and Bernard Ousmane Sane

2,3 b

Graduate School of Science and Technology, Nara Institute Science and Technology, Ikoma, Japan

Graduate School of Media and Governance, Keio University Shonan Fujisawa Campus, Kanagawa, Japan

Quantum Computing Center, Keio University, Kanagawa, Japan

Keywords:

Trusted Execution Environment, Encrypted Database, Oblivious Random Access Memory, Access Pattern

Leakage, Homomorphic Encryption, Trajectory Data.

Abstract:

The proliferation of smartphones and IoT devices has led to a rapid increase in the generation of trajectory data.

Managing this continuously generated data poses a signiﬁcant burden. To alleviate this burden, cloud databases

have become widespread, leading to increased storage of data on servers managed by other individuals and

organizations (third parties). However, if there are adversaries among these third parties, viewing the data

contents could lead to personal information leaks and privacy violations. Therefore, there are expectations

for the use of encrypted databases that allow searching and managing data while it remains encrypted (in

ciphertext form), without revealing the contents. Since data owners (clients) encrypt their data before storing

it, third parties cannot view the actual content. However, it is known that merely encrypting the data is not

sufﬁcient for security, as a vulnerability has been identiﬁed where the original data can be inferred from access

patterns to the encrypted database even without seeing the actual data content. In this paper, we propose an

anonymization method for access patterns on trajectory data in encrypted databases. For anonymization,

we apply Oblivious Random Access Memory (ORAM), which generates dummy accesses alongside data

aggregation and updates to make the original accesses unidentiﬁable. Trajectory data is often aggregated and

updated on a trajectory basis rather than by individual points. Therefore, directly generating dummy accesses

at the point level using ORAM leads to overhead in encrypted memory. In our proposed method, we separate

the data storage memory into upper and lower levels to make access patterns unidentiﬁable at the trajectory

level rather than the point level. The lower memory contains single points, while the upper memory contains

multiple points (capable of representing part or all of a trajectory), and dummy accesses are generated using

ORAM to make upper memory accesses mutually unidentiﬁable.

1 INTRODUCTION

Trajectory data records chronological changes in

location information, such as people’s movement

routes, vehicle operation records, and logistics track-

ing. This data is utilized in various ﬁelds including

urban planning, trafﬁc optimization, marketing analy-

sis, and understanding behavioral patterns. However,

since it contains large amounts of position coordinates

and timestamps continuously collected from sensors

like GPS, the data volume tends to become enor-

mous, making on-premises management challenging.

Cloud database services have become widespread due

https://orcid.org/0000-0003-2144-4949

https://orcid.org/0000-0002-9249-8285

to beneﬁts such as reduced server management bur-

den, lower implementation and operational costs, and

centralized data management. As a result, there has

been an increase in storing data on servers managed

by other individuals and organizations (third parties).

However, if there are adversaries among these

third parties, viewing the data contents could lead

to personal information leaks and privacy violations.

Therefore, data owners (clients) have increasingly be-

gun encrypting their data before entrusting it to third-

party servers such as cloud database services. This

allows data to be managed on third-party servers with-

out revealing its contents (keeping it in encrypted

form), and this is called an encrypted database. Un-

like encryption through database functionality itself,

the data is encrypted on the client side, thus pro-

Sasada, T. and Sane, B. O.

Memory-Saving Oblivious RAM for Trajectory Data via Hierarchical Generation of Dummy Access over Untrusted Cloud Environment.

DOI: 10.5220/0013370100003899

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Information Systems Security and Privacy (ICISSP 2025) - Volume 2, pages 635-642

ISBN: 978-989-758-735-1; ISSN: 2184-4356

635

tecting privacy even if the database administrator has

malicious intent. But multiple studies have revealed

that encrypted database alone does not provide suf-

ﬁcient protection. A particularly signiﬁcant vulner-

ability is information leakage through access pattern

disclosure. Access patterns are chronological records

of data access logs, and statistical analysis of these

patterns can enable inference of conﬁdential informa-

tion about the original encrypted data. In fact, (Islam

et al., 2012) demonstrated that 80% of search queries

could be inferred by analyzing access patterns to en-

crypted email repositories, highlighting the severity

of this vulnerability.

To address this vulnerability, access pattern con-

cealment methods using Oblivious Random Access

Memory (ORAM) (Goldreich and Ostrovsky, 1996)

with homomorphic encryption have been proposed.

ORAM is a technology that makes access patterns

unidentiﬁable by generating dummy accesses in ad-

dition to regular accesses. The good compatibility

between ORAM and homomorphic encryption pri-

marily stems from the ability to eliminate noise in

statistical measurements caused by dummy accesses.

In conventional ORAM, while dummy accesses are

performed to conceal access patterns, these dummy

accesses would create noise in statistical measure-

ments. By using homomorphic encryption, the con-

tent of dummy accesses can be set to encrypted zeros

(or other neutral values). This allows for increasing

the number of accesses to hide access patterns while

simultaneously preventing any impact on statistical

measurements (Moataz et al., 2015; Liu et al., 2018;

Falk et al., 2023).

In this paper, we propose Memory-saving

ORAM for trajectory Data Over Encrypted Database,

MORADO. While we apply ORAM for anonymiza-

tion, trajectory data is often aggregated and updated

on a trajectory basis rather than by individual points.

Therefore, directly generating dummy accesses at the

point level using ORAM leads to overhead in the en-

crypted database. In our proposed method, we sep-

arate the ORAM memory that stores data into up-

per and lower levels to make access patterns uniden-

tiﬁable at the trajectory level rather than the point

level. The ORAM lower memory contains single

points, while the ORAM upper memory contains mul-

tiple points (capable of representing part or all of a

trajectory), and dummy accesses are generated us-

ing ORAM to make upper memory accesses mutually

unidentiﬁable. This enables anonymization of access

patterns at the trajectory level. The structure of this

paper is as follows: In Section 2, we explain the basic

technologies such as ORAM, TDX in this research.

In Section 3, we explore the detailed design of pro-

posed method and ﬂow of dummy access generation.

In Section 4, we present the results of the experi-

mental evaluation. In Section 2, we outline related

works on databases utilizing homomophirc encryp-

tion, TEE, and their challenges. Finally, in Section

6, we conclude the paper by summarizing the contri-

butions of this research.

2 PRELIMINARY

2.1 Partially Homomorphic

Cryptosystem

A Partially Homomorphic Cryptosystem is a special

encryption scheme that allows operations (particu-

larly addition) between ciphertexts. When the result

of these operations is decrypted, it matches the re-

sult of performing the same operations on the original

plaintexts. One representative additive homomorphic

encryption scheme is the Paillier Homomorphic En-

cryption (PHE).

The key generation in PHE proceeds as follows.

First, select two large prime numbers p and q, and

calculate their product n and the least common mul-

tiple λ of p − 1 and q − 1. Next, choose an integer g

from Z

∗

and calculate the multiplicative inverse µ of

L(g mod n

) modulo n, where:

L(u) =

u −1

(1)

The resulting public key is (n, g) and the private key is

(λ, µ). In the encryption process, a random value r is

selected from Z

∗

, and for a plaintext m, the ciphertext

is computed as:

c = g

· r

mod n

(2)

For decryption, given a ciphertext c, the plaintext is

recovered by computing:

m = L(c

mod n

) · µ mod n (3)

An important point is that while both encryption and

decryption use modular exponentiation, decryption

has a higher computational cost. This is because the

base of exponentiation during decryption is the ci-

phertext. The size of ciphertext is signiﬁcantly larger

than the public key n. An important property of PHE

is that when the product of two ciphertexts is de-

crypted, it yields the sum of the corresponding plain-

texts. That is:

Dec(Enc(m

) · Enc(m

) mod n

) = m

+ m

mod n

(4)

This property enables additive operations to be per-

formed on encrypted data.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

636

2.2 Trust Domain Extensions

Intel Trust Domain Extensions (TDX) is a TEE based

on secure virtualization (Intel, 2023). TDX has the

capability to deploy hardware isolated virtual ma-

chines called trust domains (TD). TDs are isolated

from virtual machine manager (VMM)/hypervisor

and any other software which are not related to TD as

shown. This strong isolation provides the required se-

curity guarantees for a TEE system. Moreover, Intel

TDX uses multi-key total memory encryption (MK-

TME) and hashing to maintain the conﬁdentiality and

integrity of the code and data in TD. Intel TDX mod-

ule is designed to run in Secure Arbitration Mode

(SEAM). SEAM introduces an expansion to the Vir-

tual Machine Extension (VMX) architecture, estab-

lishing a fresh VMX root mode known as SEAM

root. This SEAM root mode serves as the platform for

accommodating a CPU-attested module designed for

generating VM guests termed TD. SEAM mode can

be used as two logical modes: TDX non-root mode

and TDX root mode. TDX root mode is used for

Host side operations and non-root mode is used for

TD guest operations.

2.3 Oblivious Random Access Memory

ORAM (Oblivious RAM) (Goldreich and Ostrovsky,

1996) is a primitive that obscures a user’s (proces-

sor’s) access patterns to storage (DRAM). ORAM

transforms a user’s sequence of program address ac-

cesses into access patterns that appear random. While

physical access locations remain visible to an at-

tacker, the ORAM interface guarantees that physi-

cal access patterns are independent of logical access

patterns, preventing leakage of data-dependent access

patterns. Additionally, it uses probabilistic encryption

to protect data content and hide whether updates have

occurred.

Path ORAM (Stefanov et al., 2018) is currently

the most efﬁcient and well-researched ORAM imple-

mentation. It consists of two main hardware compo-

nents: binary tree storage and an ORAM controller.

Each node in the binary tree can store up to Z useful

data blocks, with dummy blocks ﬁlling empty slots.

The ORAM controller is trusted hardware that in-

cludes a position map and stash. In Path ORAM,

accessing a data block involves consulting the posi-

tion map, reading and decrypting blocks along the

path, remapping to a new random position, and writ-

ing blocks back from the stash. This ensures each

ORAM access is random and untraceable. How-

ever, Path ORAM incurs signiﬁcant energy and per-

formance penalties compared to regular DRAM.

Write-only ORAM (WoORAM) (Li and Datta,

2017) is a lightweight version of ORAM that only

obscures write access patterns. It offers better perfor-

mance than full ORAM against attackers who cannot

monitor read access patterns. Li and Datta’s scheme

was proposed for private information retrieval in data

centers but is not efﬁcient in the context of secure pro-

cessors.

DetWoORAM (Deterministic Write-only ORAM)

(Roche et al., 2017) is a deterministic Write-only

ORAM scheme that doesn’t require a stash. It gener-

ates ﬁxed, deterministic physical write access patterns

regardless of logical write access patterns. While this

eliminates the need for a stash, it requires additional

accesses to move data from temporary storage to per-

sistent main memory, resulting in performance penal-

ties. Hardware implementation complexity also re-

mains a challenge.

3 PROPOSED METHOD

3.1 Threat Model

For the architectural design of the proposed method,

we deﬁne a threat model centered on protecting data

and access patterns. This model serves as the founda-

tion for protecting sensitive data and its usage patterns

in cloud environments. In the threat model, we con-

sider the Cloud Service Provider (CSP) as the primary

threat actor. We assume the CSP has extensive access

rights across all software layers (OS, hypervisor, and

other system components). These access rights po-

tentially enable the CSP to observe and analyze the

original data and its processing. Among the CSP’s

attack capabilities, we particularly focus on memory

dump attacks and cold boot attacks (Yitbarek et al.,

2017; Halderman et al., 2009). In memory dump at-

tacks, the CSP can acquire and analyze system mem-

ory contents at any time, risking exposure of unen-

crypted data and access patterns. In cold boot attacks,

original data might be extracted from residual RAM

data by acquiring memory contents immediately after

system power-down. Furthermore, CSP can contin-

uously monitor and view data loaded into memory.

This capability enables the CSP to track data ﬂows

and processing patterns in detail and analyze tempo-

ral correlations. This suggests the possibility of in-

ferring not only data contents but also access patterns

and DO’s behavior patterns.

In addition to CSP-related threats, malware and

malicious software targeting the cloud must also be

considered. These can lead to the theft of sensi-

tive information such as usernames and passwords,

Memory-Saving Oblivious RAM for Trajectory Data via Hierarchical Generation of Dummy Access over Untrusted Cloud Environment

637

as well as entire databases hosted in clusters. This

means that beyond data interception and tampering,

there are threats of access to pattern observation and

recording. Speciﬁc threats include malware such as

Siloscape, which targets Windows containers in cloud

environments, and Kinsing Malware, which targets

Docker/Kubernetes clusters. These malware speci-

mens can remain dormant in systems for extended pe-

riods, secretly monitoring data ﬂows and access pat-

terns. As a summary, the main protection targets in

this model are the conﬁdentiality of DO’s data and

the privacy of its access patterns. In addition to pro-

tecting data content, information about who accesses

the data, when and how is treated as equally important

protection targets.

3.2 Overall Architecture

The concept of the proposed method is shown in

the Figure 1. The proposed method leverages the

fact that, in the context of trajectory data insertion,

the majority of queries are trajectory-unit rather than

point-unit queries, thereby reducing the total amount

of dummy accesses generated by ORAM. The archi-

tecture of the proposed method is broadly divided

between inside and outside the cloud. Within the

cloud, memory is divided into several upper and lower

memories. Each upper memory consists of mul-

tiple smaller lower memories. DO encrypts their

trajectory data using Paillier homomorphic encryp-

tion (PHE) and sends it to the cloud side through a

TLS/SSL connection. Homomorphically encrypted

data insertions from DOs are ﬁrst aggregated into ap-

propriate lower memories within these upper memo-

ries. Each upper memory is assigned its own ORAM

client, with corresponding ORAM servers existing on

the database server. On the proxy server, to conceal

the distribution of requests, the number of requests

processed by each upper memory is uniformized.

Speciﬁcally, a value λ ∗ (u, α) is calculated, repre-

senting the maximum number of requests in all up-

per memories. Each upper memory appears to pro-

cess λ ∗ (u, α) requests, even when the actual num-

ber of requests is lower. On the database server

side, each ORAM server stores homomorphically en-

crypted data (

∑

Enc(d1), ...,

∑

Enc(dλ∗(u, α))) from

the corresponding upper memory. This prevents the

CSP from knowing how many requests each upper

memory is processing.

3.3 Flow of Dummy Access Generation

As input, the algorithm accepts a set E = ε

, ..., ε

consisting of w homomorphically encrypted values.

Figure 1: MORADO’s System Model Architecture.

Each encrypted value ε

is deﬁned as ε

= Enc(r

where Enc is an encryption function applied to

the corresponding original request r

. Additionally,

the algorithm requires a target lower memory set

lower

= m

lower

, ..., m

lower

and seed values S = ε

, ε

for dummy access generation as input parameters.

The expected output is an indistinguishable assign-

ment of these requests to the target lower memorys.

Phase 1. Access Aggregation: The ﬁrst phase exe-

cutes system-wide aggregation of encrypted requests.

Speciﬁcally, the cumulative value Σε for each lower

memory within all upper memory m

upper

is initialized

to zero. Subsequently, for each lower memory m

lower

(1 ≤ j ≤ α), its upper memory m

upper

is identiﬁed,

and the encrypted value ε

is added to the cumulative

value Σε of lower memory m

lower

Phase 2. Dummy Access Calculation: The second

phase determines the number of dummy accesss re-

quired for privacy protection. For each upper mem-

ory m

upper

, the number of non-zero lower memorys λ

is calculated. Here, λ

represents the crucial metric

of how many lower memorys within m

upper

contain

actual requests. Subsequently, the maximum value

∗

= max

1≤i≤u

across all upper memorys is de-

rived. This λ

∗

functions as a fundamental parameter

deﬁning the system-wide privacy protection level and

plays a decisive role in the subsequent dummy access

generation process.

Phase 3. Request Processing: The third phase

implements a two-stage reﬁned processing for

each upper memory m

upper

. The ﬁrst stage fo-

cuses on real request processing: for each lower

memory m

lower

holding aggregated requests Σε

an aggregate request(Σε

, m

lower

) is issued to

ORAM client i. This processing enables secure

transfer of real requests. The second stage handles

dummy access generation and issuance, processing

each index j from λ

+ 1 to λ

∗

. During this stage,

a lower memory m

lower

is randomly selected, and

an XOR operation ε

= ε

⊕ ε

is applied. The re-

sulting value is transmitted to ORAM client i via

aggregate request(ε

, m

lower

). Finally, either ε

or ε

in seed value S is probabilistically replaced with

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

638

Algorithm 1: MORADO.

Input: Set of encrypted values

E = {ε

, ..., ε

} where ε

= Enc(r

)

Target lower memory set

lower

= {p

, ..., p

}

Seed values S = {ε

, ε

}

Output: Indistinguishable m

lower

1 Phase 1: access aggregation

2 foreach upper memory m

upper

3 foreach lower memory m

lower

in m

upper

4 Σε ← 0

5 end

6 end

7 for j ← 1 to α do

8 Identify m

upper

containing m

lower

9 Σε

lower

← Σε

lower

+ ε

10 end

11 Phase 2: dummy access Calculation

12 foreach upper memory m

upper

13 Calculate λ

← number of non-zero lower

memorys in m

upper

14 end

15 λ

∗

← max

1≤i≤u

16 Phase 3: Request Processing

17 foreach upper memory m

upper

// Stage 1: Process real

requests

18 foreach lower memory m

lower

with

Σε

̸= 0 in m

upper

19 aggregate request(Σε

, m

lower

) to

ORAM client i

20 end

// Stage 2: Generate and

process dummy accesss

21 for j ← λ

+ 1 to λ

∗

22 Randomly select lower memory

lower

in m

upper

23 ε

← ε

⊕ ε

24 aggregate request(ε

, m

lower

) to

ORAM client i

25 Replace either ε

or ε

with ε

in S

Randomly.

26 end

27 end

3.4 Security Analysis

We prove the security of the MORADO in the fol-

lowing mathematical indistinguishability. In Morado,

an access request R

access

in the form (id, Enc(s), c)

is submitted from DOs. Let D

view

represent the

state (view) of the database prior to R

access

, where

view

consists of the aggregate values in all lower

memories. The proxy server generates the updated

view D

′

view

after R

access

. Then, CSP (adversary) A

receive both D

view

and D

′

view

with id and Enc(s).

Then, the proxy server gets the corresponding lower

memory m

lower

= Π(c), and chooses another c

′

̸= c

such that m

lower’

= Π(c

′

) ̸= m

lower

randomly. Both

lower

and m

lower’

are submitted to A . The adver-

sary guesses which of m

lower

or m

lower’

corresponds

to access R

access

. The mathematical indistinguishabil-

ity requires that for any probabilistic polynomial-time

adversary A for all possible m, N, and R

access

Pr[A(V

, V

′

, id, Enc(s),m

lower

, m

lower’

) = m

lower

]

(5)

≤

+ ε(δ) (6)

In this Equation, ε(·) is a mask mechanism and δ

is the security parameter. The adversary downloads

view D

view

from the proxy server. In MORADO,

view

contains a sequence of ORAM accesss, and

contain the DO’s ID. Lower memories is in the form

(id

upper

, id

lower

), where id

upper

and id

lower

represent

the upper memory ID and lower memory ID within

the upper memory. After downloading D

view

, the ad-

versary sends the ID to the proxy server, and the proxy

server returns the correct lower memory m

lower

(id

upper

, id

lower

) and a false lower memory m

lower’

(id

′

upper

, id

′

lower

) for the adversary to estimate between

lower

and m

lower’

. In the worst case, the adversary

is enable to match the correct access ID j within the

batch with ID, but the adversary cannot distinguish

between id

upper

and id

′

upper

. The adversary also is not

able to distinguish between id

lower

and id

′

lower

, this is

based on the mathematical security provided by the

ORAM. Then, ORAM determines whether the access

sequences having id

lower

and id

′

lower

are the same or

different. It means, the adversary cannot distinguish

between m

lower

and m

lower’

, and MORADO guaran-

tees indistinguishability.

4 EVALUATION

This experiment use a 2016 ride record dataset ob-

tained and processed from the NYC Taxi and Limou-

sine Commission and synthetic data. This dataset

contains approximately 2.08 million records, with

each record consisting of pickup timestamps and GPS

location information for taxi rides. By dividing the

Memory-Saving Oblivious RAM for Trajectory Data via Hierarchical Generation of Dummy Access over Untrusted Cloud Environment

639

Figure 2: Execution Time.

latitude and longitude points into sufﬁciently small,

equal-sized waypoints grids and excluding inacces-

sible areas, we ultimately obtained 16,367 location

grids.

4.1 Execution Time

As the Figure 2 showing, the proposed method sig-

niﬁcantly faster than DetWoORAM, which is faster

than Path ORAM. In the synthetic dataset, while Path

ORAM takes approximately 5118.53 seconds to exe-

cute, DetWoORAM requires only 18.05 seconds, and

the proposed method achieves an even shorter time of

4.11 seconds. Similar trends are observed with real-

world datasets, where Path ORAM takes 10237.06

seconds, compared to DetWoORAM’s 13.65 sec-

onds and the proposed method’s 3.05 seconds, show-

ing substantial improvements. Overall, the proposed

method achieves a 51.39% reduction compared to

using DetWoORAM alone. This is because the

DetWoORAM-only approach requires more decryp-

tion operations for each access and experiences re-

dundant updates. What’s particularly noteworthy is

that in both datasets, the improvement from Path

ORAM to DetWoORAM is dramatic (orders of mag-

nitude), and the optimization from DetWoORAM to

the proposed method achieves about a 4x performance

improvement. This demonstrates the high practicality

of the proposed method.

The left boxplot in Figure 3 evaluates the scalabil-

ity of the MORADO framework with respect to the

total number of M

lower

. The graph measures the pro-

cessing time for 1000 accesses while varying M

lower

from 2

to 2

. The right boxplot in Figure 3 evalu-

ates the system’s scalability when varying the total of

accesses α. It examines how processing time changes

as the number of accesses increases from 0 to 10000,

and the graph shows that the increase in processing

time gradually becomes more moderate as the number

of accesses grows. A notable point is that even with

the current implementation, 10000 accesses can be

processed within about 17.5 seconds, suggesting that

Figure 3: Overhead of access Processing.

further performance improvements can be expected

through parallelization of database query processing.

4.2 Storage Cost

The analysis demonstrates both the bandwidth

charges incurred between TEE and database, and

the primary storage costs for each component, based

on 1,000 access operations. From the graph, we

can see that for TEE storage costs, Path ORAM is

about 35 units and DetWoORAM is about 32 units,

while MORADO shows a slight increase to about 45

units. On the other hand, for database storage costs,

while Path ORAM is very high at about 4,000 units,

DetWoORAM is about 1,500 units, and MORADO

achieves a signiﬁcant reduction to about 1,200 units.

The baseline proposed method without redundant up-

dates exceeds both DetWoORAM and Path ORAM in

terms of bandwidth (included storage costs). The pro-

posed method shows increased storage costs on TEE

primarily due to dummy accesses used to hide access

distribution between SPs. However, since the pro-

posed method is based on homomorphic encryption, it

saves network bandwidth during transfers. A partic-

ularly noteworthy point is that the proposed method

(MORADO) achieves approximately 70% reduction

in database storage costs compared to Path ORAM.

This represents a signiﬁcant advantage in building

practical systems.

5 RELATED WORK

There are many approaches to realizing privacy-

preserving data access. In this section, we explain

different approaches to this problem with and without

TEEs.

In threat models where database administrator

(DBA) are considered adversaries, homomorphic en-

cryption is known as a promising approach. Homo-

morphic encryption is a cryptographic method that

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

640

Figure 4: Storage Cost.

enables computations on encrypted data, allowing for

aggregation and statistical processing without decryp-

tion. In particular, encrypted databases utilizing ho-

momorphic encryption can respond to queries while

keeping the data encrypted (Bian et al., 2023; Poddar

et al., 2016; Popa et al., 2011). However, as pointed

out by Isram et al. (Islam et al., 2012), encrypted

databases can still leak information to DBAs through

access patterns during memory operations. In en-

crypted database searches, the client sends encrypted

search keywords to the server, and the server per-

forms the search using these encrypted values. While

the documents themselves are encrypted, the mapping

of which encrypted keywords are contained in which

documents must be stored on the server side to enable

searching. As a result, the server can learn patterns

about which documents are returned when searching

with particular encrypted keywords.

In threat models where CSP are considered adver-

saries, TEE approaches have garnered attention. TEE

is a technology that provides a trusted execution envi-

ronment, with hardware-based implementations like

Intel SGX being widely used. Using TEE enables

isolation of programs and data running on the cloud,

even from CSP. There is also substantial research

on implementing databases within TEE-generated en-

crypted memory (enclaves) rather than using homo-

morphic encryption (Yang et al., 2024; Suzuki et al.,

2024; Yoshimura et al., 2023; Vinayagamurthy et al.,

2019). However, like homomorphic encryption-based

encrypted databases, information can still leak to the

CSP through access patterns during memory deploy-

ment. TEE (such as Intel TDX or AMD SEV-SNP)

executes protected virtual machines or processes, but

memory management (paging) must depend on the

host OS. This is due to TEE physical memory ca-

pacity limitations, the need for efﬁcient memory man-

agement, and integration with OS-level resource man-

agement. Consequently, information about which

memory pages are accessed, when page faults oc-

cur, and page replacement patterns is exposed, allow-

ing attackers to infer program execution ﬂow, observe

memory access patterns, and collect information at

the cache line level. As a result, control ﬂows like

conditional branches and data-dependent memory ac-

cess patterns are exposed, enabling attackers to in-

fer program behavior and data characteristics through

side-channel attacks, potentially leaking privacy in-

formation. Neither HE nor TEE methods can pre-

vent access pattern leakage. This is why ORAM is

being researched as a technique to conceal access

patterns. There are various oblivious protocols, in-

cluding those that make data packets unidentiﬁable.

Notable examples include Oblivious Transfer(Rabin,

2005), which allows a receiver to obtain speciﬁc in-

formation from multiple pieces of information held

by the sender without the sender knowing which in-

formation was selected. It can be used for decoupling

statistic from data volume(Sasada et al., 2023; Sasada

et al., 2022). Oblivious Message Routing(Kirman and

Martinez, 2010), which prevents information leakage

through communication pattern analysis by conceal-

ing message sources and destinations along the com-

munication path.

6 CONCLUSION

This research paper presents a novel approach to

protecting trajectory data in cloud computing envi-

ronments through a memory-saving ORAM imple-

mentation. The proposed method successfully ad-

dresses the challenges of securing sensitive location

data while maintaining system efﬁciency. By imple-

menting a hierarchical memory structure that man-

ages both trajectory-level and point-level access pat-

terns, the solution achieves signiﬁcant performance

improvements over existing approaches. The exper-

imental results demonstrate reduction in storage cost

compared to PathORAM, while maintaining robust

security measures against potential threats from CSP.

However, our security model has vulunerability in

collusion attack between DOs and CSP. First, gen-

erate security parameter δ pseudonymous IDs in our

system. Then, during a short time period id

upper

, emu-

late access from these DOs regarding the target lower

memory m

lower

. If the victim DO id sends a access

during the same period, the colluding CSP succeed in

determining whether m

lower

is the target lower mem-

ory of id’s access by validating whether the total of

ORAM accesses in this update round is δ or δ + 1.

Our future work is prevention from this attack by in-

creasing the cost of making pseudonymous IDs or by

mitigating attack via increasing the number of autho-

rized DOs.

Memory-Saving Oblivious RAM for Trajectory Data via Hierarchical Generation of Dummy Access over Untrusted Cloud Environment

641

ACKNOWLEDGEMENTS

This research was partly supported by JSPS KAK-

ENHI Grant Numbers JP22J23910 and the Daiichi-

Sankyo ”Habataku” Support Program for the Next

Generation of Researchers, NAIST Senju Monju

Project.

REFERENCES

Bian, S., Zhang, Z., Pan, H., Mao, R., Zhao, Z., Jin, Y.,

and Guan, Z. (2023). He3db: An Efﬁcient and Elastic

Encrypted Database via Arithmetic-And-Logic Fully

Homomorphic Encryption. In Proceedings of the

2023 ACM SIGSAC Conference on Computer and

Communications Security, pages 2930–2944.

Falk, B. H., Nema, R., and Ostrovsky, R. (2023). Linear-

Time 2-Party Secure Merge from Additively Homo-

morphic Encryption. Journal of Computer and System

Sciences, 137:37–49.

Goldreich, O. and Ostrovsky, R. (1996). Software Protec-

tion and Simulation on Oblivious RAMs. Journal of

the ACM, 43(3):431–473.

Halderman, J. A., Schoen, S. D., Heninger, N., Clarkson,

W., Paul, W., Calandrino, J. A., Feldman, A. J., Appel-

baum, J., and Felten, E. W. (2009). Lest We Remem-

ber: Cold-Boot Attacks on Encryption Keys. Commu-

nications of the ACM, 52(5):91–98.

Intel (2023). Intel Trust Domain Extensions (Intel

TDX) Module v1.5 Base Architecture Speciﬁcation.

https://www.intel.com/content/www/us/en/developer/

articles/technical/inteltrust-domain-extensions.html.

Islam, M. S., Kuzu, M., and Kantarcioglu, M. (2012).

Access Pattern Disclosure on Searchable Encryption:

Ramiﬁcation, Attack and Mitigation. In Network and

Distributed System Security Symposium, volume 20,

page 12. Citeseer.

Kirman, N. and Martinez, J. F. (2010). A Power-Efﬁcient

All-Optical On-Chip Interconnect Using Wavelength-

Based Oblivious Routing. ACM Sigplan Notices,

45(3):15–28.

Li, L. and Datta, A. (2017). Write-Only Oblivious RAM-

Based Privacy-Preserved Access of Outsourced Data.

International Journal of Information Security, 16:23–

42.

Liu, Z., Huang, Y., Li, J., Cheng, X., and Shen, C. (2018).

DivORAM: Towards a Practical Oblivious RAM with

Variable Block Size. Information Sciences, 447:1–11.

Moataz, T., Mayberry, T., and Blass, E.-O. (2015). Constant

Communication ORAM with Small Blocksize. In

Proceedings of the 22nd ACM SIGSAC Conference on

Computer and Communications Security, pages 862–

873.

Poddar, R., Boelter, T., and Popa, R. A. (2016). Arx: A

Strongly Encrypted Database System. IACR Cryptol.

ePrint Arch., 2016:591.

Popa, R. A., Redﬁeld, C. M., Zeldovich, N., and Balakr-

ishnan, H. (2011). CryptDB: Protecting Conﬁdential-

ity with Encrypted Query Processing. In Proceedings

of the twenty-third ACM symposium on operating sys-

tems principles, pages 85–100.

Rabin, M. O. (2005). How to Exchange Secrets with

Oblivious Transfer. IACR Cryptology ePrint Archive,

2005(187).

Roche, D. S., Aviv, A., Choi, S. G., and Mayberry,

T. (2017). Deterministic, Stash-Free Write-Only

ORAM. In Proceedings of the 2017 ACM SIGSAC

Conference on Computer and Communications Secu-

rity, pages 507–521.

Sasada, T., Taenaka, Y., and Kadobayashi, Y. (2022).

Decoupling Statistical Trends from Data Volume on

LDP-Based Spatio-Temporal Data Collection. In

2022 IEEE Future Networks World Forum, pages

262–269. IEEE.

Sasada, T., Taenaka, Y., and Kadobayashi, Y. (2023). Obliv-

ious Statistic Collection with Local Differential Pri-

vacy in Mutual Distrust. IEEE Access, 11:21374–

21386.

Stefanov, E., Dijk, M. v., Shi, E., Chan, T.-H. H., Fletcher,

C., Ren, L., Yu, X., and Devadas, S. (2018). Path

ORAM: An Extremely Simple Oblivious RAM Pro-

tocol. Journal of the ACM, 65(4):1–26.

Suzuki, T., Sasada, T., Taenaka, Y., and Kadobayashi,

Y. (2024). Mosaicdb: An efﬁcient trusted/untrusted

memory management for location data in database.

The Sixteenth International Conference on Advances

in Databases, Knowledge, and Data Applications,

pages 1–6.

Vinayagamurthy, D., Gribov, A., and Gorbunov, S. (2019).

Stealthdb: a scalable encrypted database with full sql

query support. Proceedings on Privacy Enhancing

Technologies.

Yang, X., Yue, C., Zhang, W., Liu, Y., Ooi, B. C., and Chen,

J. (2024). Secudb: An in-enclave privacy-preserving

and tamper-resistant relational database. Proceedings

of the VLDB Endowment, 17(12):3906–3919.

Yitbarek, S. F., Aga, M. T., Das, R., and Austin, T. (2017).

Cold Boot Attacks Are still Hot: Security Analysis of

Memory Scramblers in Modern Processors. In 2017

IEEE International Symposium on High Performance

Computer Architecture, pages 313–324. IEEE.

Yoshimura, M., Sasada, T., Taenaka, Y., and Kadobayashi,

Y. (2023). Memory efﬁcient data-protection for

database utilizing secure/unsecured area of intel sgx.

The Sixteenth International Conference on Advances

in Databases, Knowledge, and Data Applications,

pages 45–50.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

642