Mobile AIGC Gets a Security Upgrade with ProSecutor

Table of Links

Abstract and 1. Introduction

1.1 Background

1.2 Motivation

1.3 Our Work and Contributions and 1.4 Organization

2. Related Work

   2.1 Mobile AIGC and Its QoE Modeling

   2.2 Blockchain for Mobile Networks

3. Preliminaries

4. Prosecutor Design

   4.1 Architecture Overview

   4.2 Reputation Roll-up

   4.3 Duplex Transfer Channel

5. OS2a: Objective Service Assessment for Mobile AIGC

   5.1 Inspiration from DCM

   5.2 Objective Quality of the Service Process

   5.3 Subjective Experience of AIGC Outputs

6. OS2A on Prosecutor: Two-Phase Interaction for Mobile AIGC

   6.1 MASP Selection by Reputation

   6.2 Contract Theoretic Payment Scheme

7. Implementation and Evaluation

   7.1 Implementation and Experimental Setup

   7.2 Prosecutor Performance Evaluation

   7.3 Investigation of Functional Goals

   7.4 Security Analysis

8. Conclusion and References

Abstract—Mobile AI-Generated Content (AIGC) has achieved great attention in unleashing the power of generative AI and scaling the AIGC services. By employing numerous Mobile AIGC Service Providers (MASPs), ubiquitous and low-latency AIGC services for clients can be realized. Nonetheless, the interactions between clients and MASPs in public mobile networks, pertaining to three key mechanisms, namely MASP selection, payment scheme, and fee-ownership transfer, are unprotected. In this paper, we design the above mechanisms in a systematic approach and present the first blockchain to protect mobile AIGC, called ProSecutor. Specifically, by roll-up and layer-2 channels, ProSecutor forms a two-layer architecture, realizing tamper-proof data recording and atomic feeownership transfer with high resource efficiency. Then, we present the Objective-Subjective Service Assessment (OS2A) framework, which effectively evaluates the AIGC services by fusing the objective service quality with the reputation-based subjective experience of the service outcome (i.e., AIGC outputs). Deploying OS2A on ProSecutor, firstly, the MASP selection can be realized by sorting the reputation. Afterward, the contract theory is adopted to optimize the payment scheme and help clients avoid moral hazards in mobile networks. We implement the prototype of ProSecutor on BlockEmulator. Extensive experiments demonstrate that ProSecutor achieves 12.5× throughput and saves 67.5% storage resources compared with BlockEmulator. Moreover, the effectiveness and efficiency of the proposed mechanisms are validated.

1 INTRODUCTION

SPARKED by the phenomenal success of ChatGPT, AIGenerated Content (AIGC) has attracted significant attention from both industry and academia [1]. As the latest paradigm for content creation in the Metaverse era, AIGC enables computers to generate multimedia outputs automatically (e.g., images, videos, even 3D avatars), significantly promoting generation efficiency and saving massive time and cost. Moreover, it also makes professional artwork creation accessible for even untrained users and stimulates people’s creativity. From 2022 to 2023, we have witnessed the successful attempt of AIGC in various fields, such as Stable Diffusion in text-to-image generation, ChatGPT in Q & A, and Microsoft Copilot in the daily office. According to Acumen, AIGC is projected to achieve a global market size of USD 110.8 Billion by 2030, growing at a compound annual growth rate of 34.3% from 2022 to 2030 [2].

1.1 Background

With the deepening of AIGC applications, the scalability concern is eminently exposed. Currently, most AIGC services rely on large pre-trained models with billions of parameters, consuming considerable storage and computation resources. For instance, running Stable Diffusion requires at least one NVIDIA Ampere GPU with 6 GB memory [1], which is unaffordable for many resource-constrained clients [3]. To this end, researchers recently presented the concept of Mobile AIGC and successfully developed a series of ondevice AIGC models, e.g., MediaPipe and PaLM 2-Gecko by Google [4] and the lightweight Stable Diffusion by Chen et al. [5]. In the mobile AIGC era, clients can request AIGC inferences from Mobile AIGC Service Providers (MASPs) [6]. Since MASPs are close to clients, low service latency can be realized. Additionally, clients are able to customize the AIGC services, e.g., sharing real-time background information with MASPs to render immersive 3D environments. Furthermore, the network-wide resources and service requests can be provisioned, forming the AIGC-as-a-Service paradigm [7]. Despite these advantages, the interactions between clients and MASPs in mobile AIGC are complicated, pertaining to the following mechanisms.

• MASP Selection: In mobile AIGC, each client can access multiple nearby MASPs with varying computing power, capability, and reliability. Hence, the MASP selection mechanism should incorporate these factors and select the best MASP with the highest probability of meeting the client’s service requirements.

• Payment Scheme: Afterward, the client confirms service details with the selected MASP. A payment scheme is required, which specifies the payment method (e.g., pre-paid or post-paid) and the amount of the service fee (e.g., fixed or floating value) according to the service quality promised by the MASP.

• Fee-Ownership Transfer: Once finishing the AIGC inferences, a transfer mechanism should be employed. In this way, the client and MASP can transfer the service fee and the ownership of the AIGC output to each other in a secure manner.

1.2 Motivation

Although similar mechanisms have been studied separately in many other scenarios, the unique features of mobile AIGC bring brand-new challenges. Firstly, in traditional service markets, such as edge offloading, effective service provider selection can be realized by firstly modeling the Quality of Experience (QoE) from the client perspective and then selecting the service providers leading to the highest QoE [7], [8]. However, such schemes fail to support the emerging AIGC scenario due to the following reasons.

• Multimodality: AIGC is going beyond multimedia content generation and aiming to provide an immersive fusion of multimodal services [6]. However, most quantitative metrics for QoE measurement are modalityspecific [9], [10], [11]. Hence, we need to extend various QoE models that adapt to different AIGC modalities, which are inflexible and cannot support the evercomplicated mobile AIGC applications.

• Subjectivity: AIGC outputs can be regarded as novel digital artwork whose judgment suffers from intrinsic subjectivity. Different clients may evaluate an AIGC output from different aspects. For instance, even if an AIGC image performs well in the PyTorch Image Quality tests [12], it may not achieve satisfying QoE if its style (e.g., realism or abstractism) does not match the client’s expectations and personal preference.

For the payment scheme, the clients suffer from information asymmetry in mobile AIGC [13]. To be specific, since the resources invested by MASPs for performing AIGC inferences are unobserved, the clients are threatened by the moral hazard [14]. In this case, if the clients pay the fixed AIGC service fee in one lump sum, dishonest MASPs might not provide high-quality service as promised to save computation resources. Finally, the fee-ownership transfers in mobile AIGC are vulnerable since the anonymous clients and MASPs may repudiate without being afraid of prosecution. For example, clients can cancel ongoing payments immediately after receiving the AIGC output and vice versa. Consequently, the atomicity, i.e., whether the operations in one transfer all occur or nothing occurs, is broken.