There’s a related research idea called “moving target defense”. MT6D is similar to what you described:
MT6D: A Moving Target IPv6 Defense
https://security.vt.edu/docs/MT6D_A_Moving_Target_IPv6_Defense.pdf
The Internet Protocol version 6 (IPv6) brings with it a seemingly endless supply of network addresses. It does not, however, solve many of the vulnerabilities that existed in Internet Protocol version 4 (IPv4). In fact, privacy-related crimes in IPv6 are made easier due to the way IPv6 addresses are formed. We developed a Moving Target IPv6 Defense (MT6D) that leverages the immense address space of IPv6. The two goals of MT6D are maintaining user privacy and protecting against targeted network attacks. These goals are achieved by repeatedly rotating the addresses of both the sender and receiver. Address rotation occurs, regardless of the state of ongoing sessions, to prevent an attacker from discovering the identities of the two communicating hosts. Rotating addresses mid-session prevents an attacker from even determining that the same two hosts are communicating. The continuously changing addresses also force an attacker to repeatedly reacquire the target node before he or she can launch a successful network attack. Our proof of concept demonstrates the feasibility of MT6D and its ability to seamlessly bind new IPv6 addresses. We also demonstrate MT6D’s ability to rotate addresses mid-session without dropping or renegotiating sessions. Since MT6D operates at the network layer of the protocol stack, it provides a powerful moving target solution that is both platform and application independent.
Related, there’s a paper that uses mobile IPv6 and contrasts itself to MT6D:
Scalable Anti-Censorship Framework Using Moving Target Defense for Web Servers
https://censorbib.nymity.ch/pdf/Heydari2017a.pdf
Although the Internet has become a hub around which every aspect of our lives-from commerce to leisurely activities-is centered, many around the world are not able to freely access information over the Internet. Some governments censor what the people can and cannot see. In this paper, regardless of the socio-political view points, we focus on the design of anti-censorship technology that can be implemented on the side of the information purveyors. The primary objective is to develop a framework for combating censorship. Our approach aims to make it too expensive and impractical for the adversary to censor Web sites. In particular, we propose the use of Mobile IPv6 to form a moving target defense strategy, where the Web servers logically behave as if they are the mobile nodes (without actually moving). The potential efficacy of this framework is modeled analytically. Probabilistic models are used to derive important metrics and parameters. One key factor termed swarming ratio enables hosting sites to reason about the amount of resources needed to force the adversary’s costs over practical limits. This model is used to guide the performance goals and architectural setup of the prototype implementation (modifications are made on the server-side software and Kernel without changing the standard Mobile IPv6 protocol). Hence, the solution can be utilized without any changes to the existing network infrastructure. Furthermore, we introduce a novel, credit-based accounting strategy for grouping of users to drastically shift resource requirements in our favor. Lab-based tests are used to measure performance overheads, and based on the findings, targeted optimizations are performed to consider practical deployment scenarios. The end result is a solution that may also be combined with existing anti-censorship methods (that are end-user-based and/or assisted by friendly network assets) to form a robust anti-censorship solution.
You can also compare to GoHop, which doesn’t use multiple IP addresses, but multiple ports on the same server:
GoHop: Personal VPN to Defend from Censorship
https://censorbib.nymity.ch/pdf/Wang2014a.pdf
https://github.com/bigeagle/gohop
One way to achieve the goal of splitting a stream up and distributing it over multiple nodes is to use a Turbo Tunnel design.