Does a Larger Privacy Zone Protect You on Strava?

So far I've written two posts covering reverse-engineering Strava's privacy zones, and I found something unsettling: Strava's privacy zones can be reverse-engineered with relative ease and from just a few activities. Once you understand how it works the "hidden" start point isn't very hidden at all.

Strava lets you customise the size of the privacy zone, so the natural next question is: Does making the privacy zone larger actually help?

Strava lets you expand the radius up to 1600m, which is a useful privacy feature. The trade-off is largely cosmetic: in the worst instance a 5km run loses a few kilometres on the public map, making it look like you teleported in and out. All the stats are still there though, and the extra radius should make reverse-engineering harder.

The real question is how much harder? Does doubling or quadrupling the radius meaningfully improve privacy? Let's find out.

Is Bigger Better (For Privacy Zones)?

Setup

To test this, I reused the same 40 activities from the earlier experiments. Each activity started at the same start location in the city, and everything else was left identical. The only variable I changed was the privacy zone radius:

• 200m - the default baseline
• 400m - double the default
• 800m - quadruple the default

I didn't go all the way to 1600m because many of the activities I generated are too short for this; a 1600m radius would swallow both start and end points entirely, leaving almost nothing to show on the map. Unfortunately, Strava doesn't let you configure different start and end radii, and this can only be adjusted manually. If you wanted to test a radius up to 1600m you'd need activities that are ~5km long to be safe.

With each radius configured, I re-uploaded all activities via the Strava API, letting it apply the new privacy zones.

This is my baseline from post I: 40 activities, all clipped by a 200m privacy zone. The circle is already pretty obvious - and I've shown this can be reverse-engineered reasonably well (down to ~30m).

At 400m, everything spreads out more. The routes sit noticeably farther from the true start point and you can feel the radius doing something useful. But the shape is still unmistakable: the activity start points trace a circle. You don't need sophisticated tools to guess what sits in the middle.

At 800m, the routes push even further out. This is where things start to look a little bit safer - the circle is larger, the activities are more dispersed, and the true start feels buried somewhere in that much larger ring.

But the key issue persists: the geometry of the privacy zone is still a circle. Even if it's much wider, it still points back toward a single centre.

At first glance, this seems harder to reverse-engineer - but is it meaningfully so?

Now here's all of them combined - 200m, 400m and 800m. The larger zones clearly dwarf the smaller ones, but no matter how big they are, every privacy zone still shares the same geometry: a circle.

So visually, bigger zones help - but whether they actually protect you better is what I need to test next.

Results

With the setup done, I can now run the same geometric methods from my two earlier posts and see how they perform for the different sizes of privacy zones.

First, a reminder of the 200m case.

All methods do well here. The best prediction (the Donut Overlap Method) lands just 28.5m from the true start location, and the rest are all within 40m.

At 400m, the cloud of start points spreads and average distance from the true start increases - but the picture is nuanced.

The best method here (best polygon centre) lands at 24.8m, which is actually better than that method's result at 200m. That's not a contradiction - it just shows that single method performance can bounce around depending on start point geometry.

The Simple Average Method, however, is worse: 77.4m. That increase is mostly driven by a few outliers - notably that one start point up to the top-right - which skews the average prediction from the true start, even though some methods (like best polygon centre) still locate the centre accurately. In short: some methods (robust geometric fits) still find the centre well; naïve averages suffer when a few starts wander out of this pattern.

At 800m, distance from true start increases significantly.

The best polygon centre method comes in at 126.7m from the true start location, with the best circle centre essentially identical - both landing ~125m down the road. The remaining methods fare a bit worse, drifting a few blocks away. This is a clear improvement in privacy compared with the 200m and 400m cases: once the radius reaches 800m, even the strongest geometric methods see distances from the true start increase beyond 100m, meaning the larger radius now provides a more meaningful protective effect.

Unlike the 400m results - where a single outlier heavily shaped the average - the 800m picture is less about one anomalous point and more about the overall geometry of the clipped routes. At 800m, clipped points spread farther and vary more in direction, weakening the tight geometric pattern seen at smaller radii. Algorithms still find the general centre, but distances from the true start now exceed 100m - marking the first radius where the privacy zone delivers a clearly stronger effect.

To understand what's going on, let's take a closer look at a few of the individual methods.

The Boundary Method produces a reasonable estimate: ~125m from the true start.

The predicted centre sits directly south of the true start location. Interestingly the top of the polygon flattens out instead of stretching upward, which happens because several of the northernmost points are closer to the real start than the southern ones - so the geometry subtly pulls the estimate downward.

This downward bias comes from the 10 northernmost points averaging 611m north of the true start and the 10 southernmost points averaging 836m south of the true start. That extra southern distance pulls nearly all methods downward, amplified by slight randomness in Strava's clipping at larger radii.

Top 10 Northern Points
Average Latitudinal Distance From True Start Location: 611.5m
Median Latitudinal Distance From True Start Location: 648.2m
Range Latitudinal Distance From True Start Location: 473.3-799.0m

Top 10 Southern Points
Average Latitudinal Distance From True Start Location: 836.5m
Median Latitudinal Distance From True Start Location: 883.5m
Range Latitudinal Distance From True Start Location: 650.8-938.5m

Why the asymmetry? It's hard to say. It may simply be quirks in how Strava calculates very large privacy zones - possibly even slight elliptical distortion. Something to investigate further.

Some of this could come from Strava's own algorithm. Privacy-zone clipping isn't a perfect hard cutoff; there's slight randomness or jitter, and at larger radii this variation becomes more noticeable.

The best circle centre method behaves almost identically to the Boundary Method, coming in less than a metre away.

The circle isn't perfectly uniform - with the northernmost points again lying closer to the true start and the southernmost points stretching farther from the true start, pulling the estimated centre south.

The Donut Overlap Method still produces a reasonable estimate, but the “hot” zone is very small - which pushes the prediction too far south.

A real attacker could simply relax their clustering thresholds, widening the hot zone and nudging the prediction closer to the Boundary and Circle Fit Methods estimates.

The Adaptive Donuts Method lands further out at 188.6m.

The hot zone is large (and includes the true start), but because the southern points extend so much farther, the cluster centre shifts downward again.

Interestingly, this heatmap implies you could reduce the privacy-zone radius from 800m to ~400m (roughly 367-460m) greatly reducing the search space for an attacker - a 74% reduction in area (2.01 km² → 0.52 km²).

Key Patterns Across All Results

The table summarizes all methods across the three privacy zone radii, revealing a clear trend: as the radius increases, predicted start locations generally move farther from the true start. The relationship isn't perfectly proportional, but the trend is evident. The jump from 400m to 800m is particularly noticeable, showing that larger privacy zones provide a meaningful boost to privacy. (Note: this is based on a single test set, so further validation is needed.)

The chart visualizes the trend for each method. You can see a clear uptick in distance from the true start as the radius increases, with the biggest jumps appearing between 400m and 800m.

The takeaway is simple: increasing the privacy zone radius generally makes it harder to reverse-engineer the true start location (using these reasonably simple geometric methods), offering stronger protection than the default 200m radius. The effect is most pronounced at 800m, where even the most precise methods struggle to get very close to the true start location. I suspect the trend would be similar for 1600m.

Wrapping Up - What This Means (And What's Next)

So, does bigger mean better? Mostly, yes. At 400m, the best methods still land within ~50m of the true start; at 800m, this grows to ~125m. That's a noticeable improvement over the default 200m, but it's not bulletproof - clever methods can still get reasonably close.

These results show that larger privacy zones do make reverse-engineering harder, but they're not a perfect shield. And remember, this is all in a city environment - how would things change in suburbs or rural areas? That's exactly what I'll explore next time in Privacy Zones Fail Everywhere on Strava, But in Different Ways.