Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

SVTPrivacy: update member variable names #3255

Merged
merged 1 commit into from
Feb 25, 2025
Merged

SVTPrivacy: update member variable names #3255

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 9 additions & 9 deletions nvflare/app_common/filters/svt_privacy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,10 +45,10 @@ def __init__(

super().__init__(supported_data_kinds=[DataKind.WEIGHTS, DataKind.WEIGHT_DIFF], data_kinds_to_filter=data_kinds)

self.frac = fraction # fraction of the model to upload
self.eps_1 = epsilon
self.fraction = fraction # fraction of the model to upload
self.epsilon = epsilon
self.eps_2 = None # to be derived from eps_1
self.eps_3 = noise_var
self.noise_var = noise_var
self.gamma = gamma
self.tau = tau
self.replace = replace
Expand Down Expand Up @@ -76,21 +76,21 @@ def process_dxo(self, dxo: DXO, shareable: Shareable, fl_ctx: FLContext) -> Unio
)

# precompute thresholds
n_upload = np.minimum(np.ceil(np.float64(delta_w.size) * self.frac), np.float64(delta_w.size))
n_upload = np.minimum(np.ceil(np.float64(delta_w.size) * self.fraction), np.float64(delta_w.size))

# eps_1: threshold with noise
lambda_rho = self.gamma * 2.0 / self.eps_1
lambda_rho = self.gamma * 2.0 / self.epsilon
threshold = self.tau + np.random.laplace(scale=lambda_rho)
# eps_2: query with noise
self.eps_2 = self.eps_1 * (2.0 * n_upload) ** (2.0 / 3.0)
self.eps_2 = self.epsilon * (2.0 * n_upload) ** (2.0 / 3.0)
lambda_nu = self.gamma * 4.0 * n_upload / self.eps_2
self.logger.info(
"total params: %s, epsilon: %s, "
"perparam budget %s, threshold tau: %s + f(eps_1) = %s, "
"clip gamma: %s",
delta_w.size,
self.eps_1,
self.eps_1 / n_upload,
self.epsilon,
self.epsilon / n_upload,
self.tau,
threshold,
self.gamma,
Expand All @@ -107,7 +107,7 @@ def process_dxo(self, dxo: DXO, shareable: Shareable, fl_ctx: FLContext) -> Unio
self.log_info(fl_ctx, "selected {} responses, requested {}".format(len(accepted), n_upload))
accepted = np.random.choice(accepted, size=np.int64(n_upload), replace=self.replace)
# eps_3 return with noise
noise = np.random.laplace(scale=self.gamma * 2.0 / self.eps_3, size=accepted.shape)
noise = np.random.laplace(scale=self.gamma * 2.0 / self.noise_var, size=accepted.shape)
self.log_info(fl_ctx, "noise max: {}, median {}".format(np.max(np.abs(noise)), np.median(np.abs(noise))))
delta_w[accepted] = np.clip(delta_w[accepted] + noise, a_min=-self.gamma, a_max=self.gamma)
candidate_idx = list(set(np.arange(delta_w.size)) - set(accepted))
Expand Down
Loading