Periodic Reporting for period 4 - CARENET (Content-Aware Wireless Networks: Fundamental Limits, Algorithms, and Architectures)
Reporting period: 2023-04-01 to 2024-09-30
The ERC AdV Project CARENET focuses on massive on-demand content data distribution based on coded caching.
The project has a strong information theoretic and coding theoretic focus, and considers the fundamental limits of on-demand data delivery through a communication network, where the data is stored in one or more servers, and must be delivered on demand to multiple users via a communication network with given topology.
Since the data is already available (e.g. a library of multimedia content files), and the users have locally accessible cache memory, then the data can be pre-stored strategically in the network during off-peak time in order to facilitate the delivery at peak time. This approach, known as ``caching'', is well known and widely applied in data distribution networks. However, unlike conventional content distribution networks, ``coded'' caching exploit the power of network coding and has the potential of achieving much better scaling of the network load versus the number of demanding users.
To give a concrete and simple example: consider a system with one wireless server and two users. Users 1 demands file A, but has cached file B. User 2 demands file B, but has file A in its cache.
Then the served can broadcast the XOR A+B of the two files, satisfying both demands with a single (coded) file transmission. On the other hand, conventional caching without coding would need to transmit both files A and B,
thus doubling the load on the broadcast downlink. Generalizing this idea, coded caching schemes are network coding schemes that can exploit coding at network nodes and cache content as ``side information'' to increase the opportunity
that any given transmitted message is simultaneously useful for multiple users. This sort of multiplication effect of the utility of transmissions is referred to as ``coded caching gain''.
It turns out that the concept of coded caching can be applied in full generality to a variety of network topologies, and the investigation of the theoretical limits, the practical coding algorithms, and the demonstration and implementation on an actual wireless network, are the subject of the research carried out in CARENET.
The societal impact of this research consists of the improvement of communication networks and in particular wireless networks, which has demonstrated a fully critical and very important infrastructure especially in the last year and a half of Covid pandemic. If the economy and society in EU and in the world did not reach a complete halt and collapse it is (mainly) thanks to the fact that citizens have widespread access to broadband internet and massive segments of society have been transferred from the real world (office, shops, factories) to the virtual online world. The pandemic should be a reminder of the fact that an efficient and reliable access to internet is a critical strategic factor to guarantee society and economy resilience. Therefore, improving this critical infrastructure has a clear societal impact in improving the readiness and resilience level of our society.
The project has a strong information theoretic and coding theoretic focus, and considers the fundamental limits of on-demand data delivery through a communication network, where the data is stored in one or more servers, and must be delivered on demand to multiple users via a communication network with given topology.
Since the data is already available (e.g. a library of multimedia content files), and the users have locally accessible cache memory, then the data can be pre-stored strategically in the network during off-peak time in order to facilitate the delivery at peak time. This approach, known as ``caching'', is well known and widely applied in data distribution networks. However, unlike conventional content distribution networks, ``coded'' caching exploit the power of network coding and has the potential of achieving much better scaling of the network load versus the number of demanding users.
To give a concrete and simple example: consider a system with one wireless server and two users. Users 1 demands file A, but has cached file B. User 2 demands file B, but has file A in its cache.
Then the served can broadcast the XOR A+B of the two files, satisfying both demands with a single (coded) file transmission. On the other hand, conventional caching without coding would need to transmit both files A and B,
thus doubling the load on the broadcast downlink. Generalizing this idea, coded caching schemes are network coding schemes that can exploit coding at network nodes and cache content as ``side information'' to increase the opportunity
that any given transmitted message is simultaneously useful for multiple users. This sort of multiplication effect of the utility of transmissions is referred to as ``coded caching gain''.
It turns out that the concept of coded caching can be applied in full generality to a variety of network topologies, and the investigation of the theoretical limits, the practical coding algorithms, and the demonstration and implementation on an actual wireless network, are the subject of the research carried out in CARENET.
The societal impact of this research consists of the improvement of communication networks and in particular wireless networks, which has demonstrated a fully critical and very important infrastructure especially in the last year and a half of Covid pandemic. If the economy and society in EU and in the world did not reach a complete halt and collapse it is (mainly) thanks to the fact that citizens have widespread access to broadband internet and massive segments of society have been transferred from the real world (office, shops, factories) to the virtual online world. The pandemic should be a reminder of the fact that an efficient and reliable access to internet is a critical strategic factor to guarantee society and economy resilience. Therefore, improving this critical infrastructure has a clear societal impact in improving the readiness and resilience level of our society.
The main results achieved in the CARENET project are:
1) We have fully characterized the fundamental limits of Device-to-Device (D2D) coded caching, establishing conditions for the exact optimality of an improved scheme with respect to our previously proposed scheme (which was order-optimal but not exactly optimal) under uncoded prefetching (i.e. under the condition that the users store directly segments of the library files, and not functions thereof).
2) We have formulated a new problem of coded caching with privacy, where the privacy of the users demands must be preserved with respect to the other users in the system. In fact, in the original schemes of coded caching, every user could be aware and easily determine which file other users have demanded. In contrast, we have proposed a novel scheme to achieve information theoretic privacy of the demands and characterized its order-optimality, both for the single server topology and for the D2D topology.
3) We have significantly expanded the network topologies for which coded caching was previously studied, including novel ``Fog'' topologies with partially shared caches, coded caching over MIMO channels, multi-access coded caching for shared caches under combinatorial topologies, hierarchical networks, and two-dimensional distributed caching networks.
4) We have considered the implementation of coded caching over practical ``routing networks'', that serves as a guideline for implementing coded caching at the application layer, i.e. ``above IP''. In particular, these were: the inability of intermediate nodes such as wireless routers and base stations to store cached content; the fact that coded caching should be run at the server-client level, by some third party content distribution entity, while the core network and the wireless access are fixed by some protocol (e.g. 5G) and run by another entity (a wireless network provider); the problem of subpacketization, which must be limited for finite length files; the problem of decentralized operations, for which each user should be able to cache independently from other users; the problem of asynchronous streaming sessions, where each user starts and stop at arbitrary times.
5) We have fully characterized the problem of coded caching under different channel quality and video quality to the users. In particular, we have laid the foundations for
variable adaptive quality coded caching schemes that can work in synergy with the standard quality adaptation schemes used by state-of-the art streaming services such as YouTube and Netflix,
based on the so-called DASH approach (dynamic adaptive streaming over HTTP).
6) We have extensively tested our ideas over a physical platform for coded caching over WiFi networks built at TU Berlin in collaboration with CADAMI, a startup company based in Munich.
7) We have extended the coded caching scenarios from file delivery to more general computation tasks, i.e. the delivery of functions of data files.
We have determined the fundamental information theoretic limits of coded caching for linear function computation, where each user demands an arbitrary linear combination of the data files.
We have also considered the problem of data shuffling, where users must exchange data such that each user, at the end of the exchange, has a different data block in an assigned data block permutation.
Furthermore, we have exploited the formal similarity and common coding approaches to address problems in secure data aggregation in federated learning, which is an important paradigm of deep learning on large distributed data sets involving proprietary user data. One of the most exciting outcomes of CARENET is the discovery of the many formal analogies and coding technique similarities between a number of apparently unrelated problems, such as coded caching, private information retrieval, coded distributed computing, and secure federated learning. All together, the output of the project yields a new fundamental and unifying framework to these many problems in network coding and information theory.
1) We have fully characterized the fundamental limits of Device-to-Device (D2D) coded caching, establishing conditions for the exact optimality of an improved scheme with respect to our previously proposed scheme (which was order-optimal but not exactly optimal) under uncoded prefetching (i.e. under the condition that the users store directly segments of the library files, and not functions thereof).
2) We have formulated a new problem of coded caching with privacy, where the privacy of the users demands must be preserved with respect to the other users in the system. In fact, in the original schemes of coded caching, every user could be aware and easily determine which file other users have demanded. In contrast, we have proposed a novel scheme to achieve information theoretic privacy of the demands and characterized its order-optimality, both for the single server topology and for the D2D topology.
3) We have significantly expanded the network topologies for which coded caching was previously studied, including novel ``Fog'' topologies with partially shared caches, coded caching over MIMO channels, multi-access coded caching for shared caches under combinatorial topologies, hierarchical networks, and two-dimensional distributed caching networks.
4) We have considered the implementation of coded caching over practical ``routing networks'', that serves as a guideline for implementing coded caching at the application layer, i.e. ``above IP''. In particular, these were: the inability of intermediate nodes such as wireless routers and base stations to store cached content; the fact that coded caching should be run at the server-client level, by some third party content distribution entity, while the core network and the wireless access are fixed by some protocol (e.g. 5G) and run by another entity (a wireless network provider); the problem of subpacketization, which must be limited for finite length files; the problem of decentralized operations, for which each user should be able to cache independently from other users; the problem of asynchronous streaming sessions, where each user starts and stop at arbitrary times.
5) We have fully characterized the problem of coded caching under different channel quality and video quality to the users. In particular, we have laid the foundations for
variable adaptive quality coded caching schemes that can work in synergy with the standard quality adaptation schemes used by state-of-the art streaming services such as YouTube and Netflix,
based on the so-called DASH approach (dynamic adaptive streaming over HTTP).
6) We have extensively tested our ideas over a physical platform for coded caching over WiFi networks built at TU Berlin in collaboration with CADAMI, a startup company based in Munich.
7) We have extended the coded caching scenarios from file delivery to more general computation tasks, i.e. the delivery of functions of data files.
We have determined the fundamental information theoretic limits of coded caching for linear function computation, where each user demands an arbitrary linear combination of the data files.
We have also considered the problem of data shuffling, where users must exchange data such that each user, at the end of the exchange, has a different data block in an assigned data block permutation.
Furthermore, we have exploited the formal similarity and common coding approaches to address problems in secure data aggregation in federated learning, which is an important paradigm of deep learning on large distributed data sets involving proprietary user data. One of the most exciting outcomes of CARENET is the discovery of the many formal analogies and coding technique similarities between a number of apparently unrelated problems, such as coded caching, private information retrieval, coded distributed computing, and secure federated learning. All together, the output of the project yields a new fundamental and unifying framework to these many problems in network coding and information theory.
The project has pushed the knowhow in coded caching networks well beyond the state of the art of the topic at the starting date of the project (October 2018).