!19324 Add differential privacy parameters for mindspore federated learning.

Merge pull request !19324 from jxlang910/fl-dp

mindspore/ccsrc/pipeline/jit/init.cc

@@ -397,7 +397,13 @@ PYBIND11_MODULE(_c_expression, m) {
     .def("set_config_file_path", &PSContext::set_config_file_path,
          "Set configuration files required by the communication layer.")
     .def("config_file_path", &PSContext::config_file_path,
-         "Get configuration files required by the communication layer.");
+         "Get configuration files required by the communication layer.")
+    .def("set_dp_eps", &PSContext::set_dp_eps, "Set dp epsilon for federated learning secure aggregation.")
+    .def("set_dp_delta", &PSContext::set_dp_delta, "Set dp delta for federated learning secure aggregation.")
+    .def("set_dp_norm_clip", &PSContext::set_dp_norm_clip,
+         "Set dp norm clip for federated learning secure aggregation.")
+    .def("set_encrypt_type", &PSContext::set_encrypt_type,
+         "Set encrypt type for federated learning secure aggregation.");
 
   (void)py::class_<OpInfoLoaderPy, std::shared_ptr<OpInfoLoaderPy>>(m, "OpInfoLoaderPy")
     .def(py::init())

mindspore/ccsrc/ps/ps_context.cc

@@ -184,6 +184,47 @@ void PSContext::set_server_mode(const std::string &server_mode) {
 
 const std::string &PSContext::server_mode() const { return server_mode_; }
 
+void PSContext::set_encrypt_type(const std::string &encrypt_type) {
+  if (encrypt_type != kNotEncryptType && encrypt_type != kDPEncryptType && encrypt_type != kPWEncryptType) {
+    MS_LOG(EXCEPTION) << encrypt_type << " is invalid. Encrypt type must be " << kNotEncryptType << " or "
+                      << kDPEncryptType << " or " << kPWEncryptType;
+    return;
+  }
+  encrypt_type_ = encrypt_type;
+}
+const std::string &PSContext::encrypt_type() const { return encrypt_type_; }
+
+void PSContext::set_dp_eps(float dp_eps) {
+  if (dp_eps > 0) {
+    dp_eps_ = dp_eps;
+  } else {
+    MS_LOG(EXCEPTION) << dp_eps << " is invalid, dp_eps must be larger than 0.";
+    return;
+  }
+}
+
+float PSContext::dp_eps() const { return dp_eps_; }
+
+void PSContext::set_dp_delta(float dp_delta) {
+  if (dp_delta > 0 && dp_delta < 1) {
+    dp_delta_ = dp_delta;
+  } else {
+    MS_LOG(EXCEPTION) << dp_delta << " is invalid, dp_delta must be in range of (0, 1).";
+    return;
+  }
+}
+float PSContext::dp_delta() const { return dp_delta_; }
+
+void PSContext::set_dp_norm_clip(float dp_norm_clip) {
+  if (dp_norm_clip > 0) {
+    dp_norm_clip_ = dp_norm_clip;
+  } else {
+    MS_LOG(EXCEPTION) << dp_norm_clip << " is invalid, dp_norm_clip must be larger than 0.";
+    return;
+  }
+}
+float PSContext::dp_norm_clip() const { return dp_norm_clip_; }
+
 void PSContext::set_ms_role(const std::string &role) {
   if (server_mode_ != kServerModeFL && server_mode_ != kServerModeHybrid) {
     MS_LOG(EXCEPTION) << "Only federated learning supports to set role by fl context.";

mindspore/ccsrc/ps/ps_context.h

@@ -35,9 +35,9 @@ constexpr char kEnvRoleOfServer[] = "MS_SERVER";
 constexpr char kEnvRoleOfWorker[] = "MS_WORKER";
 constexpr char kEnvRoleOfScheduler[] = "MS_SCHED";
 constexpr char kEnvRoleOfNotPS[] = "MS_NOT_PS";
-constexpr char kDPEncryptType[] = "DPEncrypt";
-constexpr char kPWEncryptType[] = "PWEncrypt";
-constexpr char kNotEncryptType[] = "NotEncrypt";
+constexpr char kDPEncryptType[] = "DP_ENCRYPT";
+constexpr char kPWEncryptType[] = "PW_ENCRYPT";
+constexpr char kNotEncryptType[] = "NOT_ENCRYPT";
 
 // Use binary data to represent federated learning server's context so that we can judge which round resets the
 // iteration. From right to left, each bit stands for:
@@ -166,6 +166,18 @@ class PSContext {
   void set_config_file_path(const std::string &path);
   std::string config_file_path() const;
 
+  void set_dp_eps(float dp_eps);
+  float dp_eps() const;
+
+  void set_dp_delta(float dp_delta);
+  float dp_delta() const;
+
+  void set_dp_norm_clip(float dp_norm_clip);
+  float dp_norm_clip() const;
+
+  void set_encrypt_type(const std::string &encrypt_type);
+  const std::string &encrypt_type() const;
+
  private:
   PSContext()
       : ps_enabled_(false),
@@ -199,7 +211,11 @@ class PSContext {
         secure_aggregation_(false),
         cluster_config_(nullptr),
         scheduler_manage_port_(11202),
-        config_file_path_("") {}
+        config_file_path_(""),
+        dp_eps_(50),
+        dp_delta_(0.01),
+        dp_norm_clip_(1.0),
+        encrypt_type_(kNotEncryptType) {}
   bool ps_enabled_;
   bool is_worker_;
   bool is_pserver_;
@@ -276,6 +292,18 @@ class PSContext {
 
   // The path of the configuration file, used to configure the certification path and persistent storage type, etc.
   std::string config_file_path_;
+
+  // Epsilon budget of differential privacy mechanism. Used in federated learning for now.
+  float dp_eps_;
+
+  // Delta budget of differential privacy mechanism. Used in federated learning for now.
+  float dp_delta_;
+
+  // Norm clip factor of differential privacy mechanism. Used in federated learning for now.
+  float dp_norm_clip_;
+
+  // Secure mechanism for federated learning. Used in federated learning for now.
+  std::string encrypt_type_;
 };
 }  // namespace ps
 }  // namespace mindspore

mindspore/ccsrc/ps/server/kernel/round/reconstruct_secrets_kernel.cc

@@ -133,7 +133,7 @@ bool ReconstructSecretsKernel::Launch(const std::vector<AddressPtr> &inputs, con
 
 void ReconstructSecretsKernel::OnLastCountEvent(const std::shared_ptr<core::MessageHandler> &message) {
   MS_LOG(INFO) << "ITERATION NUMBER IS : " << LocalMetaStore::GetInstance().curr_iter_num();
-  if (true) {  // todo: PSContext::instance()->encrypt_type == PWEncrypt {
+  if (PSContext::instance()->encrypt_type() == kPWEncryptType) {
     while (!Executor::GetInstance().IsAllWeightAggregationDone()) {
       std::this_thread::sleep_for(std::chrono::milliseconds(5));
     }

mindspore/ccsrc/ps/server/kernel/round/update_model_kernel.cc

@@ -96,9 +96,10 @@ void UpdateModelKernel::OnLastCountEvent(const std::shared_ptr<core::MessageHand
     size_t total_data_size = LocalMetaStore::GetInstance().value<size_t>(kCtxFedAvgTotalDataSize);
     MS_LOG(INFO) << "Total data size for iteration " << LocalMetaStore::GetInstance().curr_iter_num() << " is "
                  << total_data_size;
-
+    if (PSContext::instance()->encrypt_type() != kPWEncryptType) {
       FinishIteration();
     }
+  }
 }
 
 bool UpdateModelKernel::ReachThresholdForUpdateModel(const std::shared_ptr<FBBuilder> &fbb) {

mindspore/ccsrc/ps/server/server.cc

@@ -235,6 +235,10 @@ void Server::InitCipher() {
   unsigned char cipher_p[SECRET_MAX_LEN] = {0};
   int cipher_g = 1;
   unsigned char cipher_prime[PRIME_MAX_LEN] = {0};
+  float dp_eps = PSContext::instance()->dp_eps();
+  float dp_delta = PSContext::instance()->dp_delta();
+  float dp_norm_clip = PSContext::instance()->dp_norm_clip();
+  std::string encrypt_type = PSContext::instance()->encrypt_type();
 
   mpz_t prim;
   mpz_init(prim);
@@ -248,10 +252,10 @@ void Server::InitCipher() {
   param.t = cipher_t;
   memcpy_s(param.p, SECRET_MAX_LEN, cipher_p, SECRET_MAX_LEN);
   memcpy_s(param.prime, PRIME_MAX_LEN, cipher_prime, PRIME_MAX_LEN);
-  // param.dp_delta = dp_delta;
-  // param.dp_eps = dp_eps;
-  // param.dp_norm_clip = dp_norm_clip;
-  param.encrypt_type = kNotEncryptType;  // PSContext::instance()->encrypt_type;
+  param.dp_delta = dp_delta;
+  param.dp_eps = dp_eps;
+  param.dp_norm_clip = dp_norm_clip;
+  param.encrypt_type = encrypt_type;
   cipher_init_->Init(param, 0, cipher_initial_client_cnt_, cipher_exchange_secrets_cnt_, cipher_share_secrets_cnt_,
                      cipher_get_clientlist_cnt_, cipher_reconstruct_secrets_down_cnt_,
                      cipher_reconstruct_secrets_up_cnt_);

mindspore/context.py

@@ -851,6 +851,17 @@ def set_fl_context(**kwargs):
         client_learning_rate (float): Client training learning rate. Default: 0.001.
         worker_step_num_per_iteration (int): The worker's standalone training step number before communicating with
                                              server. Default: 65.
+        dp_eps (float): Epsilon budget of differential privacy mechanism. The smaller the dp_eps, the better the
+            privacy protection effect. Default: 50.0.
+        dp_delta (float): Delta budget of differential privacy mechanism, which is usually equals the reciprocal of
+            client number. The smaller the dp_delta, the better the privacy protection effect. Default: 0.01.
+        dp_norm_clip (float): A factor used for clipping model's weights for differential mechanism. Its value is
+            suggested to be 0.5~2. Default: 1.0.
+        encrypt_type (string): Secure schema for federated learning, which can be 'NOT_ENCRYPT', 'DP_ENCRYPT' or
+            'PW_ENCRYPT'. If 'DP_ENCRYPT', differential privacy schema would be applied for clients and the privacy
+            protection effect would be determined by dp_eps, dp_delta and dp_norm_clip as described above. If
+            'PW_ENCRYPT', pairwise secure aggregation would be applied to protect clients' model from stealing.
+            Default: 'NOT_ENCRYPT'.
 
     Raises:
         ValueError: If input key is not the attribute in federated learning mode context.

mindspore/parallel/_ps_context.py

@@ -68,7 +68,11 @@ _set_ps_context_func_map = {
     "worker_step_num_per_iteration": ps_context().set_worker_step_num_per_iteration,
     "enable_ps_ssl": ps_context().set_enable_ssl,
     "scheduler_manage_port": ps_context().set_scheduler_manage_port,
-    "config_file_path": ps_context().set_config_file_path
+    "config_file_path": ps_context().set_config_file_path,
+    "dp_eps": ps_context().set_dp_eps,
+    "dp_delta": ps_context().set_dp_delta,
+    "dp_norm_clip": ps_context().set_dp_norm_clip,
+    "encrypt_type": ps_context().set_encrypt_type
 }
 
 _get_ps_context_func_map = {

tests/st/fl/mobile/run_mobile_server.py

@@ -38,6 +38,11 @@ parser.add_argument("--client_batch_size", type=int, default=32)
 parser.add_argument("--client_learning_rate", type=float, default=0.1)
 parser.add_argument("--local_server_num", type=int, default=-1)
 parser.add_argument("--config_file_path", type=str, default="")
+parser.add_argument("--encrypt_type", type=str, default="NotEncrypt")
+# parameters for encrypt_type='DP_ENCRYPT'
+parser.add_argument("--dp_eps", type=float, default=50.0)
+parser.add_argument("--dp_delta", type=float, default=0.01)  # 1/worker_num
+parser.add_argument("--dp_norm_clip", type=float, default=1.0)
 
 if __name__ == "__main__":
     args, _ = parser.parse_known_args()
@@ -62,6 +67,10 @@ if __name__ == "__main__":
     client_learning_rate = args.client_learning_rate
     local_server_num = args.local_server_num
     config_file_path = args.config_file_path
+    dp_eps = args.dp_eps
+    dp_delta = args.dp_delta
+    dp_norm_clip = args.dp_norm_clip
+    encrypt_type = args.encrypt_type
 
     if local_server_num == -1:
         local_server_num = server_num
@@ -95,6 +104,10 @@ if __name__ == "__main__":
         cmd_server += " --client_epoch_num=" + str(client_epoch_num)
         cmd_server += " --client_batch_size=" + str(client_batch_size)
         cmd_server += " --client_learning_rate=" + str(client_learning_rate)
+        cmd_server += " --dp_eps=" + str(dp_eps)
+        cmd_server += " --dp_delta=" + str(dp_delta)
+        cmd_server += " --dp_norm_clip=" + str(dp_norm_clip)
+        cmd_server += " --encrypt_type=" + str(encrypt_type)
         cmd_server += " > server.log 2>&1 &"
 
         import time

tests/st/fl/mobile/test_mobile_lenet.py

@@ -46,6 +46,10 @@ parser.add_argument("--client_batch_size", type=int, default=32)
 parser.add_argument("--client_learning_rate", type=float, default=0.1)
 parser.add_argument("--scheduler_manage_port", type=int, default=11202)
 parser.add_argument("--config_file_path", type=str, default="")
+# parameters for encrypt_type='DP_ENCRYPT'
+parser.add_argument("--dp_eps", type=float, default=50.0)
+parser.add_argument("--dp_delta", type=float, default=0.01)  # 1/worker_num
+parser.add_argument("--dp_norm_clip", type=float, default=1.0)
 
 args, _ = parser.parse_known_args()
 device_target = args.device_target
@@ -70,6 +74,10 @@ client_batch_size = args.client_batch_size
 client_learning_rate = args.client_learning_rate
 scheduler_manage_port = args.scheduler_manage_port
 config_file_path = args.config_file_path
+dp_eps = args.dp_eps
+dp_delta = args.dp_delta
+dp_norm_clip = args.dp_norm_clip
+encrypt_type = args.encrypt_type
 
 ctx = {
     "enable_fl": True,
@@ -93,7 +101,11 @@ ctx = {
     "client_batch_size": client_batch_size,
     "client_learning_rate": client_learning_rate,
     "scheduler_manage_port": scheduler_manage_port,
-    "config_file_path": config_file_path
+    "config_file_path": config_file_path,
+    "dp_eps": dp_eps,
+    "dp_delta": dp_delta,
+    "dp_norm_clip": dp_norm_clip,
+    "encrypt_type": encrypt_type
 }
 
 context.set_context(mode=context.GRAPH_MODE, device_target=device_target, save_graphs=False)