Comparing Buitd-in Aggregation Schemes

This example demonstrates the comparison of eight built-in aggregation schemes. We draw 100 samples from two normal distributions with different mean and co-variance. The samples are then aggregated using the built-in aggregation rules.

import matplotlib.pyplot as plt
import numpy as np
import torch

from fedlib.aggregators import (
    Clippedclustering,
    DnC,
    GeoMed,
    Mean,
    Median,
    Multikrum,
    Trimmedmean,
)

plt.rcParams["axes.linewidth"] = 1.5  # set the value globally
plt.rcParams["font.weight"] = "bold"
plt.rcParams["font.size"] = 16
plt.rcParams["axes.labelweight"] = "bold"

np.random.seed(1)
sz = 40
sample_sz = 80

mean = np.array((0, 0))
cov = [[20, 0], [0, 20]]
benign = np.random.multivariate_normal(mean, cov, 60)
mean = np.array((30, 30))
cov = [[60, 0], [0, 60]]
outliers = np.concatenate([np.random.multivariate_normal(mean, cov, 40)])
all_data = np.concatenate([benign, outliers])

all_data_tensor = torch.Tensor(np.concatenate([benign, outliers]))

all_data_tensor = [tensor for tensor in all_data_tensor]

aggs = [
    Mean(),
    Multikrum(len(outliers), 10),
    GeoMed(),
    Median(),
    DnC(num_byzantine=len(outliers)),
    Trimmedmean(num_byzantine=len(outliers)),
    Clippedclustering(),
]

# sphinx_gallery_thumbnail_number = 1
fig, axs = plt.subplots(figsize=(8, 8))

ax = axs
ax.scatter(
    benign[:, 0],
    benign[:, 1],
    s=sample_sz,
    alpha=0.6,
    color="r",
    linewidths=0.2,
    edgecolors="black",
)
ax.scatter(
    outliers[:, 0],
    outliers[:, 1],
    s=sample_sz,
    color=[0.0, 0.7, 0.0, 1.0],
    linewidths=0.2,
    edgecolors="black",
)


def plot_agg(ax, agg):
    target = agg(all_data_tensor).cpu().detach().numpy()
    ax.scatter(
        target[0],
        target[1],
        s=sz * 3,
        label=type(agg).__name__,
        linewidths=0.3,
        edgecolors="black",
    )


list(map(lambda agg: plot_agg(ax, agg), aggs))
ax.set_xticks([])
ax.set_yticks([])
ax.legend()

fig.tight_layout(pad=0.0, w_pad=0.6, h_pad=0.5)
plt.show()

In this example, the results of Mean deviated away by the outliers. All the other are inside the range of benign data.