Trusted and decentralized aggregation for federated learning
Techniques for distributed federated learning leverage a multi-layered defense strategy to provide for reduced information leakage. In lieu of aggregating model updates centrally, an aggregation function is decentralized into multiple independent and functionally-equivalent execution entities, each running within its own trusted executed environment (TEE). The TEEs enable confidential and remote-attestable federated aggregation. Preferably, each aggregator entity runs within an encrypted virtual machine that support runtime in-memory encryption. Each party remotely authenticates the TEE before participating in the training. By using multiple decentralized aggregators, parties are enabled to partition their respective model updates at model-parameter granularity, and can map single weights to a specific aggregator entity. Parties also can dynamically shuffle fragmentary model updates at each training iteration to further obfuscate the information dispatched to each aggregator execution entity. This architectural prevents the aggregator from being a single point-of-failure, and serves to protect the model even if all aggregators are compromised.
1 . A method of providing federated learning with reduced information leakage of model aggregation, comprising:
loading each aggregator execution entity included in a set of decentralized aggregator execution entities into a separate trusted execution environment that acts as a trusted intermediary for isolating each aggregator execution entity from other aggregator execution entities in the set, where each trusted execution environment utilizes a distinct encryption key to provide runtime memory encryption protection for a respective aggregator execution entity;
registering each of multiple parties with each of the aggregator execution entities included in the set of decentralized aggregator execution entities;
establishing and maintaining a secure communication channel between the aggregator execution entities for training synchronization during the federated learning;
distributing to each aggregator execution entity included in the set of decentralized aggregator execution entities an assigned portion of a partitioned model update for a machine learning model provided by the multiple parties; and
initiating execution of the set of decentralized aggregator execution entities to train the machine learning model using the assigned portions of the partitioned model update, wherein the set of decentralized aggregator execution entities engage in a number of training iterations with the multiple parties, where at each training iteration of the machine learning model, each party of the multiple parties:
obtains a latest model update from each of the set of decentralized aggregator execution entities,
produces a next model update using local training data, and
provides the next model update to each of the set of decentralized aggregator execution entities, wherein response to receiving the next model update, the set of decentralized aggregator execution entities communicate via the secure communication channel to merge the next model updates into an aggregated machine learning model, and distribute the aggregated machine learning model to each party of the multiple parties, wherein each party of the multiple parties remotely authenticates hardware for each trusted execution environment containing a respective aggregator execution entity.
2 . The method as described in claim 1 , further including attesting to an integrity of the aggregator execution entity.
3 . The method as described in claim 1 , wherein the multiple parties partition respective model updates at model-parameter granularity and map single weights to one of the set of decentralized aggregator execution entities.
4 . The method as described in claim 3 , wherein an ordering of parameter or gradient elements in at least one partition are shuffled.
5 . The method as described in claim 1 , wherein, with respect to at least one other of the set of decentralized aggregator execution entities, the aggregator execution entity executes in one of: a different machine, a different data center, a different trusted execution environment architecture, and a different geo-location.
6 . The method as described in claim 1 , wherein loading the aggregator execution entity includes injecting a unique secret into encrypted physical memory of an encrypted virtual machine that provides the trusted execution environment and using the unique secret to authenticate the aggregator execution entity at registration of a party.
7 . An apparatus, comprising:
a hardware processor; and
a computer memory holding computer program instructions executed by the hardware processor to provide federated learning with reduced information leakage of model aggregation, the computer program instructions configured to:
load each aggregator execution entity included in a set of decentralized aggregator execution entities into a separate trusted execution environment that acts as a trusted intermediary for isolating each aggregator execution entity from other aggregator execution entities in the set, where each trusted execution environment utilizes a distinct encryption key to provide runtime memory encryption protection for a respective aggregator execution entity;
register each of multiple parties with each of the aggregator execution entities included in the set of decentralized aggregator execution entities;
establish and maintain a secure communication channel between the aggregator execution entities for training synchronization during the federated learning;
distribute to each aggregator execution entity included in the set of decentralized aggregator execution entities an assigned portion of a partitioned model update for a machine learning model provided by the multiple parties; and
initiate execution of the set of decentralized aggregator execution entities to train the machine learning model using the assigned portions of the partitioned model update, wherein the set of decentralized aggregator execution entities engage in a number of training iterations with the multiple parties, where at each training iteration of the machine learning model, each party of the multiple parties:
obtains a latest model update from each of the set of decentralized aggregator execution entities,
produces a next model update using local training data, and
provides the next model update to each of the set of decentralized aggregator execution entities, wherein response to receiving the next model update, the set of decentralized aggregator execution entities communicate via the secure communication channel to merge the next model updates into an aggregated machine learning model, and distribute the aggregated machine learning model to each party of the multiple parties, wherein each party of the multiple parties remotely authenticates hardware for each trusted execution environment containing a respective aggregator execution entity.
8 . The apparatus as described in claim 7 , wherein the computer program instructions are further configured to attest to an integrity of the aggregator execution entity.
9 . The apparatus as described in claim 7 , wherein the multiple parties partition respective model updates at model-parameter granularity and map single weights to one of the set of decentralized aggregator execution entities.
10 . The apparatus as described in claim 9 , wherein an ordering of parameter or gradient elements in at least one partition are shuffled.
11 . The apparatus as described in claim 7 , wherein, with respect to at least one other of the set of decentralized aggregator execution entities, the aggregator execution entity executes in one of: a different machine, a different data center, a different trusted execution environment architecture, and a different geo-location.
12 . The apparatus as described in claim 7 , wherein the computer program instructions configured to load the aggregator execution entity further including computer program instructions configured to receive a unique secret in the trusted execution environment and to use the unique secret to authenticate the aggregator execution entity at registration of a party.
13 . A computer program product in a non-transitory computer readable medium for use in a data processing system to provide federated learning with reduced information leakage of model aggregation, the computer program product holding computer program instructions that, when executed by the data processing system, are configured to:
load each aggregator execution entity included in a set of decentralized aggregator execution entities into a separate trusted execution environment that acts as a trusted intermediary for isolating each aggregator execution entity from other aggregator execution entities in the set, where each trusted execution environment utilizes a distinct encryption key to provide runtime memory encryption protection for a respective aggregator execution entity;
register each of multiple parties with each of the aggregator execution entities included in the set of decentralized aggregator execution entities;
establish and maintain a secure communication channel between the aggregator execution entities for training synchronization during the federated learning;
distributing to each aggregator execution entity included in the set of decentralized aggregator execution entities an assigned portion of a partitioned model update for a machine learning model provided by the multiple parties; and
initiating execution of the set of decentralized aggregator execution entities to train the machine learning model using the assigned portions of the partitioned model update, wherein the set of decentralized aggregator execution entities engage in a number of training iterations with the multiple parties, where at each training iteration of the machine learning model, each party of the multiple parties:
obtains a latest model update from each of the set of decentralized aggregator execution entities,
produces a next model update using local training data, and
provides the next model update to each of the set of decentralized aggregator execution entities, wherein response to receiving the next model update, the set of decentralized aggregator execution entities communicate via the secure communication channel to merge the next model updates into an aggregated machine learning model, and distribute the aggregated machine learning model to each party of the multiple parties, wherein each party of the multiple parties remotely authenticates hardware for each trusted execution environment containing a respective aggregator execution entity.
14 . The computer program product as described in claim 13 , wherein the computer program instructions are further configured to attest to an integrity of the aggregator execution entity.
15 . The computer program product as described in claim 13 , wherein the multiple parties partition respective model updates at model-parameter granularity and map single weights to one of the set of decentralized aggregator execution entities.
16 . The computer program product as described in claim 15 , wherein an ordering of parameter or gradient elements in at least one partition are shuffled.
17 . The computer program product as described in claim 13 , wherein, with respect to at least one other of the set of decentralized aggregator execution entities, the aggregator execution entity executes in one of: a different machine, a different data center, a different trusted execution environment architecture, and a different geo-location.
18 . The computer program instructions as described in claim 13 , wherein the computer program instructions configured to load the aggregator execution entity further including computer program instructions configured to receive a unique secret in the trusted execution environment and to use the unique secret to authenticate the aggregator execution entity at registration of a party.
19 . A federated learning system that is secure against information leakage of model aggregation, comprising:
a set of trusted execution environments; and
an aggregator partitioned into a set of decentralized aggregator execution entities, with each aggregator execution entity loaded into a respective trusted execution environment that acts as a trusted intermediary for isolating each aggregator execution entity from other aggregator execution entities in the set, and each aggregator execution entity utilizes a respective distinct encryption key to provide runtime memory encryption protection;
wherein each party of multiple parties is registered with each of the aggregator execution entities included in the set of decentralized aggregator execution entities, and wherein a secure communication channel is established between the aggregator execution entities for training synchronization during the federated learning, and wherein an assigned portion of a partitioned model update for a machine learning model provided by the multiple parties is distributed to each aggregator execution entity included in the set of decentralized aggregator execution entities; and
wherein execution of the set of decentralized aggregator execution entities is initiated to train the machine learning model using the assigned portions of the partitioned model update, wherein the set of decentralized aggregator execution entities engage in a number of training iterations with the multiple parties, where at each training iteration of the machine learning model, each party of the multiple parties:
obtains a latest model update from each of the set of decentralized aggregator execution entities,
produces a next model update using local training data, and
provides the next model update to each of the set of decentralized aggregator execution entities, wherein response to receiving the next model update, the set of decentralized aggregator execution entities communicate via the secure communication channel to merge the next model updates into an aggregated machine learning model, and distribute the aggregated machine learning model to each party of the multiple parties, wherein each party of the multiple parties remotely authenticates hardware for each trusted execution environment containing a respective aggregator execution entity.
20 . The federated learning system as described in claim 19 , wherein, with respect to at least one other of the set of decentralized aggregator execution entities, an aggregator execution entity executes in one of: a different machine, a different data center, a different trusted execution environment architecture, and a different geo-location.
21 . The federated learning system as described in claim 19 , wherein the multiple parties partition respective model updates at model-parameter granularity and map single weights to one of the set of decentralized aggregator execution entities.
22 . The method as described in claim 1 , further comprising:
partitioning a central aggregator into the set of decentralized aggregator execution entities, where each decentralized aggregator in the set is functionally-equivalent to each other; and
launching a set of encrypted virtual machines, wherein each decentralized aggregator is loaded onto a respective encrypted virtual machine that provides a respective trusted execution environment for the decentralized aggregator.